audio_pipeline

2026-06-10 17:12:58 +03:00
commit 00ddac5af7
29 changed files with 3297 additions and 0 deletions
--- a/.gitignore
+++ b/.gitignore
@@ -0,0 +1 @@
+.env
--- a/README.md
+++ b/README.md
@@ -0,0 +1,661 @@
+# Audio Pipeline
+
+Пайплайн обработки аудиозаписей звонков: от появления файла на диске до классификации, анализа по промптам и публикации итогового результата в очередь `final`.
+
+Система построена на **файловом триггере** + **RabbitMQ** + **PostgreSQL** + внешних API (Nexara STT, Yandex LLM).
+
+---
+
+## Содержание
+
+1. [Общая схема](#общая-схема)
+2. [Структура проекта](#структура-проекта)
+3. [Инфраструктура](#инфраструктура)
+4. [Этапы обработки](#этапы-обработки)
+5. [Файловое хранилище](#файловое-хранилище)
+6. [RabbitMQ](#rabbitmq)
+7. [PostgreSQL](#postgresql)
+8. [Воркеры](#воркеры)
+9. [Форматы сообщений](#форматы-сообщений)
+10. [Промпты для анализа](#промпты-для-анализа)
+11. [Агрегация и очередь final](#агрегация-и-очередь-final)
+12. [Конфигурация (.env)](#конфигурация-env)
+13. [Запуск и управление](#запуск-и-управление)
+14. [Логирование](#логирование)
+15. [Обработка ошибок](#обработка-ошибок)
+16. [Переключение промптов на API](#переключение-промптов-на-api)
+17. [Что не реализовано](#что-не-реализовано)
+
+---
+
+## Общая схема
+
+```
+┌─────────────┐     ┌──────────────┐     ┌─────────────┐
+│  incoming/  │────▶│   watcher    │────▶│  RabbitMQ   │
+│  (файлы)    │     │  (сканер)    │     │ audio.new   │
+└─────────────┘     └──────────────┘     └──────┬──────┘
+                                                 │
+                    ┌──────────────┐             ▼
+                    │ processing/  │     ┌─────────────┐
+                    │  (аудио)     │◀────│ transcribe  │── Nexara API (STT)
+                    └──────────────┘     └──────┬──────┘
+                                                │
+                                    fanout transcription_done
+                                    ┌───────────┴───────────┐
+                                    ▼                       ▼
+                             ┌─────────────┐         ┌─────────────┐
+                             │   analyse   │         │   tagging   │
+                             │ Yandex LLM  │         │ Yandex LLM  │
+                             └──────┬──────┘         └──────┬──────┘
+                                    │                       │
+                                    └───────────┬───────────┘
+                                                ▼
+                                         ┌─────────────┐
+                                         │  PostgreSQL │
+                                         │   results   │
+                                         └──────┬──────┘
+                                                │
+                              оба готовы ───────┤
+                                                ▼
+                                         ┌─────────────┐
+                                         │   final     │  (RabbitMQ)
+                                         │  + удаление │
+                                         │  файла      │
+                                         └─────────────┘
+```
+
+**Ключевые принципы:**
+
+- Каждый этап — отдельный Docker-сервис (воркер).
+- Связь между этапами — через RabbitMQ (асинхронно).
+- Промежуточные и итоговые результаты LLM — в PostgreSQL.
+- `analyse` и `tagging` работают **параллельно** после транскрипции.
+- В `final` публикует тот воркер, который завершился **последним**.
+
+---
+
+## Структура проекта
+
+```
+audio-pipeline/
+├── docker-compose.yml          # инфраструктура и все сервисы
+├── .env                        # конфигурация (не в git)
+├── db/
+│   └── init.sql                # схема PostgreSQL
+├── storage/                    # аудиофайлы на хосте (монтируется в контейнеры)
+│   ├── incoming/               # сюда кладут новые файлы
+│   ├── processing/             # файлы в обработке
+│   └── failed/                 # файлы при критических ошибках watcher
+├── watcher/                    # сканер файловой системы
+│   ├── cmd/watcher/main.go
+│   └── internal/
+│       ├── config/
+│       ├── scanner/
+│       └── publisher/
+└── workers/
+    ├── transcribe/             # STT + загрузка промптов + fanout
+    │   ├── cmd/transcribe/main.go
+    │   ├── configs/prompts.json
+    │   └── internal/
+    │       ├── consumer/
+    │       ├── nexara/
+    │       ├── prompts/
+    │       └── models/
+    ├── analyse/                # анализ по промптам (Yandex LLM)
+    │   └── cmd/analyse/main.go
+    └── tagging/                # классификация диалога (Yandex LLM)
+        └── cmd/tagging/main.go
+```
+
+---
+
+## Инфраструктура
+
+| Сервис      | Контейнер   | Порты          | Назначение                    |
+|-------------|-------------|----------------|-------------------------------|
+| `rabbit`    | rabbit      | 5672, 15672    | RabbitMQ + Management UI      |
+| `postgres`  | postgres    | 5432           | Хранение результатов          |
+| `watcher`   | watcher     | —              | Мониторинг `incoming/`        |
+| `transcribe`| transcribe  | —              | Транскрипция + fanout         |
+| `analyse`   | analyse     | —              | Анализ по промптам            |
+| `tagging`   | tagging     | —              | Классификация диалога         |
+
+**RabbitMQ UI:** http://localhost:15672 (логин/пароль из `.env`)
+
+---
+
+## Этапы обработки
+
+### 1. Watcher — обнаружение файла
+
+**Триггер:** появление аудиофайла в `{STORAGE_ROOT}/incoming/`.
+
+**Алгоритм:**
+
+1. Каждые `POLL_INTERVAL` (по умолчанию 5 с) сканирует `incoming/`.
+2. Пропускает скрытые файлы (`.`) и временные (`.tmp`).
+3. Проверяет расширение: `.mp3`, `.wav`, `.m4a`, `.ogg`, `.flac`, `.webm`.
+4. Ждёт стабилизации размера файла (`STABLE_WINDOW` / `STABLE_CHECKS`) — защита от незавершённой загрузки.
+5. Генерирует ULID `task_id`.
+6. Атомарно переименовывает: `incoming/name.wav` → `processing/{task_id}.wav`.
+7. Публикует задачу в RabbitMQ (`audio_pipeline` / `audio.new`).
+
+**При ошибке публикации:** файл возвращается в `incoming/`. Если откат невозможен — перемещается в `failed/`.
+
+### 2. Transcribe — транскрипция
+
+**Вход:** очередь `transcribe.tasks`.
+
+**Алгоритм:**
+
+1. Читает аудиофайл по `file_path` из сообщения.
+2. Отправляет в **Nexara API** (Speech-to-Text).
+3. Загружает промпты (`prompts.json` или HTTP API).
+4. Формирует `TranscriptionResult` и публикует в fanout-exchange `transcription_done`.
+5. Сообщение доставляется **одновременно** в очереди `analyse` и `tagging`.
+
+### 3. Tagging — классификация диалога
+
+**Вход:** очередь `tagging`.
+
+**Алгоритм:**
+
+1. Получает транскрипцию из сообщения.
+2. Отправляет **один** запрос в **Yandex LLM** с промптом классификации (L1/L2/L3, risk_level и т.д.).
+3. Сохраняет результат в `results.tagging` (PostgreSQL).
+4. Если `results.analysis` уже заполнен — публикует в `final` и удаляет файл.
+
+### 4. Analyse — анализ по промптам
+
+**Вход:** очередь `analyse`.
+
+**Алгоритм:**
+
+1. Получает транскрипцию и массив `prompts` из сообщения.
+2. Для **каждого** промпта — отдельный запрос в **Yandex LLM** (сейчас 3 промпта: behavioral, client_data, cargo_data).
+3. Сохраняет результат в `results.analysis`, метаданные — в `results.metadata`.
+4. Если `results.tagging` уже заполнен — публикует в `final` и удаляет файл.
+
+### 5. Final — итоговое сообщение
+
+**Выход:** очередь `final` (default exchange, routing key = `final`).
+
+Публикуется **полный JSON** из таблицы `results`: транскрипция, analysis, tagging, metadata, статус, timestamps.
+
+После успешной публикации аудиофайл удаляется из `processing/`.
+
+> **Consumer для очереди `final` пока не реализован** — сообщения накапливаются в очереди до подключения обработчика.
+
+---
+
+## Файловое хранилище
+
+Путь на хосте задаётся `STORAGE_ROOT` (по умолчанию `./storage`).
+
+| Директория    | Назначение                                      |
+|---------------|-------------------------------------------------|
+| `incoming/`   | Новые файлы для обработки                       |
+| `processing/` | Файлы в работе (после claim watcher)            |
+| `failed/`     | Файлы при невосстановимых ошибках watcher       |
+
+**Жизненный цикл файла:**
+
+```
+incoming/recording.wav
+    → processing/01KTN....wav   (watcher)
+    → остаётся до завершения    (transcribe читает по пути)
+    → удаляется                 (после публикации в final)
+```
+
+Внутри контейнеров путь: `/data/storage/...` (volume mount).
+
+---
+
+## RabbitMQ
+
+### Топология
+
+```
+exchange: audio_pipeline (direct)
+    └── queue: transcribe.tasks  ←  routing key: audio.new
+
+exchange: transcription_done (fanout)
+    ├── queue: analyse
+    └── queue: tagging
+
+queue: final  (default exchange, без binding)
+```
+
+### Очереди
+
+| Очередь            | Producer   | Consumer   | Описание                    |
+|--------------------|------------|------------|-----------------------------|
+| `transcribe.tasks` | watcher    | transcribe | Новые аудиозадачи           |
+| `analyse`          | transcribe | analyse    | Результат транскрипции      |
+| `tagging`          | transcribe | tagging    | Результат транскрипции      |
+| `final`            | analyse/tagging | —     | Итоговый результат          |
+
+### Dead Letter
+
+Очередь `transcribe.tasks` настроена с DLX (`dlx`). Сообщения с `Nack(requeue=false)` уходят в dead-letter. Отдельная DLQ-очередь может потребовать дополнительной настройки.
+
+---
+
+## PostgreSQL
+
+### Таблица `results`
+
+```sql
+CREATE TABLE results (
+    task_id        TEXT PRIMARY KEY,
+    filename       TEXT,
+    transcription  TEXT,
+    analysis       JSONB,      -- результат analyse (Yandex LLM)
+    tagging        JSONB,      -- результат tagging (Yandex LLM)
+    metadata       JSONB,      -- file_path, segments, prompts, language, transcribed_at
+    status         TEXT DEFAULT 'pending',  -- pending | done
+    created_at     TIMESTAMPTZ,
+    updated_at     TIMESTAMPTZ
+);
+```
+
+### Кто что пишет
+
+| Поле            | Пишет    | Когда                          |
+|-----------------|----------|--------------------------------|
+| `tagging`       | tagging  | После классификации            |
+| `analysis`      | analyse  | После анализа                  |
+| `filename`      | оба      | При сохранении своей части     |
+| `transcription` | analyse  | При сохранении analysis        |
+| `metadata`      | analyse  | file_path, segments, prompts…  |
+| `status`        | оба      | `done` когда оба поля заполнены|
+
+### Просмотр данных
+
+```bash
+docker exec -it postgres psql -U pipeline -d pipeline
+```
+
+```sql
+SELECT task_id, filename, status, updated_at FROM results ORDER BY updated_at DESC;
+
+SELECT task_id, jsonb_pretty(analysis), jsonb_pretty(tagging)
+FROM results WHERE task_id = 'ВАШ_TASK_ID';
+```
+
+---
+
+## Воркеры
+
+### Watcher
+
+- **Язык:** Go
+- **Зависимости:** RabbitMQ
+- **Volume:** `${STORAGE_ROOT}:/data/storage`
+- **Не использует:** Postgres, Yandex, Nexara
+
+### Transcribe
+
+- **Язык:** Go
+- **API:** Nexara (STT)
+- **Промпты:** static file или HTTP
+- **Volume:** storage + `configs/prompts.json`
+
+### Tagging
+
+- **Язык:** Go
+- **API:** Yandex Cloud LLM (`YANDEX_API_KEY`, `YANDEX_MODEL`)
+- **При старте:** тестовый запрос к Yandex API (проверка подключения)
+- **Volume:** storage (для удаления файлов) + `.env` (hot-reload токена)
+
+### Analyse
+
+- **Язык:** Go
+- **API:** Yandex Cloud LLM
+- **Вызовов LLM на задачу:** по числу промптов (сейчас 3)
+- **Без повторов:** один вызов на промпт, при ошибке — сообщение отбрасывается
+- **Volume:** storage + `.env`
+
+---
+
+## Форматы сообщений
+
+### 1. Watcher → Transcribe
+
+**Exchange:** `audio_pipeline` (direct)  
+**Routing key:** `audio.new`  
+**Queue:** `transcribe.tasks`
+
+```json
+{
+  "task_id": "01KTNVA3EKW8CY2QDAYKKVZ40W",
+  "file_path": "/data/storage/processing/01KTNVA3EKW8CY2QDAYKKVZ40W.wav",
+  "filename": "1.wav",
+  "size": 1234567,
+  "created_at": 1780914907
+}
+```
+
+### 2. Transcribe → Analyse + Tagging
+
+**Exchange:** `transcription_done` (fanout)  
+**Queues:** `analyse`, `tagging` (одинаковое тело)
+
+```json
+{
+  "task_id": "01KTNVA3EKW8CY2QDAYKKVZ40W",
+  "filename": "1.wav",
+  "file_path": "/data/storage/processing/01KTNVA3EKW8CY2QDAYKKVZ40W.wav",
+  "transcription": "полный текст транскрипции...",
+  "language": "ru",
+  "segments": [
+    {"start": 0.0, "end": 27.96, "text": "..."}
+  ],
+  "prompts": [
+    {
+      "id": 1,
+      "id_section": 1,
+      "name": "behavioral",
+      "prompt": "Ты — строгий классификатор звонков...",
+      "dt_create": "2026-06-09T09:00:00"
+    }
+  ],
+  "transcribed_at": 1780914907
+}
+```
+
+### 3. Final — итоговое сообщение
+
+**Queue:** `final`
+
+```json
+{
+  "task_id": "01KTNVA3EKW8CY2QDAYKKVZ40W",
+  "filename": "1.wav",
+  "transcription": "полный текст...",
+  "analysis": {
+    "behavioral": {
+      "greeting": {"value": true, "evidence": "Здравствуйте", "confidence": 0.95},
+      "initiative": {"value": true, "evidence": "...", "confidence": 0.8},
+      "questions_check": {"value": false, "evidence": null, "confidence": 0.0},
+      "closing": {"value": true, "evidence": "всего доброго", "confidence": 0.9}
+    },
+    "client_data": { "...": "..." },
+    "cargo_data": { "...": "..." }
+  },
+  "tagging": {
+    "L1": "tracking",
+    "L2": "location_request",
+    "L3": "delay",
+    "risk_level": "medium",
+    "has_action_items": true,
+    "has_deadline": false
+  },
+  "file_path": "/data/storage/processing/01KTNVA3EKW8CY2QDAYKKVZ40W.wav",
+  "language": "ru",
+  "segments": [...],
+  "prompts": [...],
+  "transcribed_at": 1780914907,
+  "status": "done",
+  "created_at": "2026-06-09T09:00:00Z",
+  "updated_at": "2026-06-09T09:05:00Z"
+}
+```
+
+---
+
+## Промпты для анализа
+
+### Источник
+
+Файл: `workers/transcribe/configs/prompts.json`  
+Или HTTP API (см. [переключение на API](#переключение-промптов-на-api)).
+
+Transcribe загружает промпты и **вкладывает их в сообщение** для analyse. Analyse не знает, откуда они пришли.
+
+### Текущие промпты (3 штуки)
+
+| name           | Назначение                                              |
+|----------------|---------------------------------------------------------|
+| `behavioral`   | Приветствие, инициативность, вопросы, прощание          |
+| `client_data`  | Первый раз, город, тип клиента, контакты, источник      |
+| `cargo_data`   | Характер груза, параметры, стоимость                    |
+
+Каждый промпт — **полный текст инструкции** с форматом JSON-ответа. Analyse отправляет:
+
+```
+<текст промпта из конфига>
+
+=== ТРАНСКРИПЦИЯ ===
+"""
+<текст звонка>
+"""
+```
+
+Ответ LLM сохраняется целиком под ключом `name` промпта в `analysis`.
+
+### Tagging — отдельный промпт
+
+Tagging **не использует** `prompts.json`. У него встроенный промпт классификации логистических диалогов (L1/L2/L3, risk_level, has_action_items, has_deadline).
+
+---
+
+## Агрегация и очередь final
+
+Оба воркера (`analyse`, `tagging`) пишут в одну строку `results` по `task_id`.
+
+**Атомарная проверка готовности** (SQL):
+
+```sql
+UPDATE results SET ...
+RETURNING (analysis IS NOT NULL AND tagging IS NOT NULL)
+```
+
+- Если `RETURNING = false` — воркер ждёт второго.
+- Если `RETURNING = true` — этот воркер:
+  1. Читает полную строку из БД
+  2. Публикует JSON в очередь `final`
+  3. Удаляет файл из `processing/` (только пути с `/processing/`)
+
+---
+
+## Конфигурация (.env)
+
+Пример ключевых переменных:
+
+```env
+# Storage
+STORAGE_ROOT=./storage
+
+# Watcher
+POLL_INTERVAL=5s
+STABLE_WINDOW=2s
+STABLE_CHECKS=3
+
+# RabbitMQ
+RABBITMQ_URL=amqp://admin:secret123@rabbit:5672/
+RABBITMQ_EXCHANGE=audio_pipeline
+RABBITMQ_ROUTING_KEY=audio.new
+
+# Transcribe
+INPUT_QUEUE=transcribe.tasks
+OUTPUT_EXCHANGE=transcription_done
+ANALYSE_QUEUE=analyse
+TAGGING_QUEUE=tagging
+FINAL_QUEUE=final
+PREFETCH=1
+
+# Nexara (STT)
+NEXARA_BASE_URL=https://api.nexara.ru
+NEXARA_API_KEY=...
+NEXARA_MODEL=whisper-1
+NEXARA_TIMEOUT=10m
+
+# Prompts
+PROMPTS_SOURCE=static
+PROMPTS_FILE=/app/configs/prompts.json
+PROMPTS_SECTION=1
+
+# Postgres
+POSTGRES_USER=pipeline
+POSTGRES_PASSWORD=pipeline_secret
+POSTGRES_DB=pipeline
+DATABASE_URL=postgres://pipeline:pipeline_secret@postgres:5432/pipeline?sslmode=disable
+
+# Yandex LLM (tagging + analyse)
+YANDEX_API_KEY=t1....
+YANDEX_MODEL=gpt://folder_id/model_name
+YANDEX_API_URL=https://ai.api.cloud.yandex.net/v1/chat/completions
+```
+
+### Hot-reload токена Yandex
+
+Воркеры `tagging` и `analyse` монтируют `.env` как `/config/.env` и перечитывают при **каждом старте** контейнера:
+
+```bash
+docker compose restart tagging analyse
+```
+
+> `docker compose restart` не пересоздаёт контейнер, но перезапускает процесс, который читает свежий `.env`.
+
+---
+
+## Запуск и управление
+
+### Первый запуск
+
+```bash
+cd audio-pipeline
+docker compose up -d --build
+```
+
+### Полный сброс (очереди + БД)
+
+```bash
+docker compose down -v
+docker compose up -d --build
+```
+
+### Пересборка отдельных воркеров
+
+```bash
+docker compose up -d --build transcribe analyse tagging
+```
+
+### Обработка нового файла
+
+```bash
+cp recording.wav storage/incoming/
+```
+
+### Проверка статуса
+
+```bash
+docker compose ps
+docker compose logs -f watcher transcribe analyse tagging
+```
+
+### RabbitMQ — просмотр очереди final
+
+UI: http://localhost:15672 → Queues → `final`
+
+---
+
+## Логирование
+
+Все воркеры пишут **структурированные JSON-логи** в stdout.
+
+### Ключевые события
+
+| Событие                  | Воркер    | Описание                              |
+|--------------------------|-----------|---------------------------------------|
+| `claimed file`           | watcher   | Файл взят в обработку                 |
+| `transcribed`            | transcribe| STT завершён                          |
+| `llm call ok`            | analyse/tagging | Вызов Yandex API              |
+| `task complete`          | analyse/tagging | Оба результата готовы         |
+| `published final`        | analyse/tagging | Сообщение в final             |
+| `processing file deleted`  | analyse/tagging | Файл удалён из processing   |
+| `yandex api check ok`    | analyse/tagging | Проверка API при старте       |
+
+### Поиск в логах
+
+```bash
+# все LLM-вызовы
+docker compose logs analyse 2>&1 | grep '"llm call'
+
+# завершённые задачи
+docker compose logs 2>&1 | grep '"task complete"'
+
+# ошибки
+docker compose logs 2>&1 | grep '"level":"WARN"'
+```
+
+---
+
+## Обработка ошибок
+
+| Ситуация                         | Поведение                                      |
+|----------------------------------|------------------------------------------------|
+| Битое JSON в очереди             | `Nack(requeue=false)` — в DLQ                  |
+| Ошибка Nexara STT                | `Nack(requeue=false)` — в DLQ                  |
+| Ошибка Yandex LLM                | `Nack(requeue=false)` — сообщение отбрасывается |
+| Ошибка сохранения в Postgres     | `Nack(requeue=false)` — отбрасывается          |
+| Redelivered сообщение            | Пропускается (без повторного вызова LLM)       |
+| Ошибка публикации в final        | Файл **не** удаляется                          |
+| Yandex API недоступен при старте | Воркер не запускается, контейнер рестартит     |
+
+**Политика без повторов:** каждый промпт / классификация — ровно один вызов LLM. Повторные доставки RabbitMQ игнорируются.
+
+---
+
+## Переключение промптов на API
+
+Уже реализовано в transcribe. Достаточно изменить `.env`:
+
+```env
+PROMPTS_SOURCE=http
+PROMPTS_BASE_URL=https://your-api.example.com
+PROMPTS_API_KEY=your_token
+PROMPTS_SECTION=1
+```
+
+**Запрос:** `GET {PROMPTS_BASE_URL}/metrics/?id_section=1`
+
+**Ожидаемый ответ** — массив в том же формате, что `prompts.json`:
+
+```json
+[
+  {
+    "id": 1,
+    "id_section": 1,
+    "name": "behavioral",
+    "prompt": "полный текст промпта...",
+    "dt_create": "2026-06-09T09:00:00"
+  }
+]
+```
+
+Analyse менять не нужно — промпты приходят в сообщении RabbitMQ.
+
+---
+
+## Что не реализовано
+
+- **Consumer очереди `final`** — нет воркера, который читает итоговые сообщения
+- **DLQ-очередь** `transcribe.tasks.failed` — exchange `dlx` объявлен, но отдельная очередь может не быть привязана
+- **Повторная обработка** при ошибках LLM — намеренно отключена
+- **Архивация** удалённых файлов — файлы удаляются без бэкапа
+
+---
+
+## Быстрый troubleshooting
+
+| Проблема                              | Решение                                                |
+|---------------------------------------|--------------------------------------------------------|
+| Файл не обрабатывается                | Проверить `incoming/`, логи watcher                  |
+| TLS timeout к Yandex из контейнера    | MTU Docker, VPN, `docker compose restart`              |
+| Старый Yandex токен                   | Обновить `.env`, `docker compose restart tagging analyse` |
+| `context canceled` в final            | Исправлено — пересобрать analyse/tagging               |
+| `metadata` column does not exist      | `ALTER TABLE results ADD COLUMN IF NOT EXISTS metadata JSONB;` |
+| Файл остаётся в processing            | Проверить, дошли ли оба воркера до `published final`   |
--- a/db/init.sql
+++ b/db/init.sql
@@ -0,0 +1,13 @@
+CREATE TABLE IF NOT EXISTS results (
+    task_id        TEXT PRIMARY KEY,
+    filename       TEXT,
+    transcription  TEXT,
+    analysis       JSONB,
+    tagging        JSONB,
+    metadata       JSONB,
+    status         TEXT NOT NULL DEFAULT 'pending',
+    created_at     TIMESTAMPTZ NOT NULL DEFAULT now(),
+    updated_at     TIMESTAMPTZ NOT NULL DEFAULT now()
+);
+
+ALTER TABLE results ADD COLUMN IF NOT EXISTS metadata JSONB;
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -0,0 +1,100 @@
+services:
+  rabbit:
+    image: rabbitmq:3-management-alpine
+    container_name: rabbit
+    ports:
+      - "5672:5672"
+      - "15672:15672"
+    environment:
+      RABBITMQ_DEFAULT_USER: ${RABBITMQ_DEFAULT_USER:-admin}
+      RABBITMQ_DEFAULT_PASS: ${RABBITMQ_DEFAULT_PASS:-secret123}
+    healthcheck:
+      test: ["CMD", "rabbitmq-diagnostics", "-q", "ping"]
+      interval: 10s
+      timeout: 5s
+      retries: 5
+    restart: unless-stopped
+
+  postgres:
+    image: postgres:16-alpine
+    container_name: postgres
+    ports:
+      - "5432:5432"
+    environment:
+      POSTGRES_USER: ${POSTGRES_USER:-pipeline}
+      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:-pipeline_secret}
+      POSTGRES_DB: ${POSTGRES_DB:-pipeline}
+    volumes:
+      - postgres_data:/var/lib/postgresql/data
+      - ./db/init.sql:/docker-entrypoint-initdb.d/init.sql:ro
+    healthcheck:
+      test: ["CMD-SHELL", "pg_isready -U ${POSTGRES_USER:-pipeline} -d ${POSTGRES_DB:-pipeline}"]
+      interval: 5s
+      timeout: 5s
+      retries: 10
+      start_period: 15s
+    restart: unless-stopped
+
+  watcher:
+    build: ./watcher
+    container_name: watcher
+    env_file: .env
+    volumes:
+      - ${STORAGE_ROOT:-./storage}:/data/storage
+    environment:
+      STORAGE_ROOT: /data/storage
+    depends_on:
+      rabbit:
+        condition: service_healthy
+    restart: unless-stopped
+
+  transcribe:
+    build: ./workers/transcribe
+    container_name: transcribe
+    env_file: .env
+    volumes:
+      - ${STORAGE_ROOT:-./storage}:/data/storage
+      - ./workers/transcribe/configs:/app/configs:ro
+    environment:
+      STORAGE_ROOT: /data/storage
+    depends_on:
+      rabbit:
+        condition: service_healthy
+    restart: unless-stopped
+
+  tagging:
+    build: ./workers/tagging
+    container_name: tagging
+    env_file: .env
+    volumes:
+      - ${STORAGE_ROOT:-./storage}:/data/storage
+      - ./.env:/config/.env:ro
+    environment:
+      STORAGE_ROOT: /data/storage
+      DOTENV_PATH: /config/.env
+    depends_on:
+      rabbit:
+        condition: service_healthy
+      postgres:
+        condition: service_healthy
+    restart: unless-stopped
+
+  analyse:
+    build: ./workers/analyse
+    container_name: analyse
+    env_file: .env
+    volumes:
+      - ${STORAGE_ROOT:-./storage}:/data/storage
+      - ./.env:/config/.env:ro
+    environment:
+      STORAGE_ROOT: /data/storage
+      DOTENV_PATH: /config/.env
+    depends_on:
+      rabbit:
+        condition: service_healthy
+      postgres:
+        condition: service_healthy
+    restart: unless-stopped
+
+volumes:
+  postgres_data:
--- a/watcher/Dockerfile
+++ b/watcher/Dockerfile
@@ -0,0 +1,11 @@
+FROM golang:1.22-alpine AS build
+WORKDIR /src
+COPY go.mod go.sum* ./
+RUN go mod download
+COPY . .
+RUN CGO_ENABLED=0 go build -o /watcher ./cmd/watcher
+
+FROM alpine:3.19
+RUN apk add --no-cache ca-certificates
+COPY --from=build /watcher /watcher
+ENTRYPOINT ["/watcher"]
--- a/watcher/cmd/watcher/main.go
+++ b/watcher/cmd/watcher/main.go
@@ -0,0 +1,115 @@
+package main
+
+import (
+	"context"
+	"log/slog"
+	"os"
+	"os/signal"
+	"syscall"
+	"time"
+
+	amqp "github.com/rabbitmq/amqp091-go"
+
+	"github.com/yourorg/watcher/internal/config"
+	"github.com/yourorg/watcher/internal/publisher"
+	"github.com/yourorg/watcher/internal/scanner"
+)
+
+func main() {
+	slog.SetDefault(slog.New(slog.NewJSONHandler(os.Stdout, &slog.HandlerOptions{Level: slog.LevelInfo})))
+
+	cfg := config.Load()
+	sc := scanner.New(scanner.Config{
+		StorageRoot:   cfg.StorageRoot,
+		IncomingDir:   cfg.IncomingDir,
+		ProcessingDir: cfg.ProcessingDir,
+		FailedDir:     cfg.FailedDir,
+		StableWindow:  cfg.StableWindow,
+		StableChecks:  cfg.StableChecks,
+	})
+	if err := sc.EnsureDirs(); err != nil {
+		slog.Error("ensure dirs failed", "error", err)
+		os.Exit(1)
+	}
+
+	ch := mustRabbit(cfg.RabbitURL)
+	if err := ch.ExchangeDeclare(cfg.Exchange, "direct", true, false, false, false, nil); err != nil {
+		slog.Error("declare exchange failed", "error", err)
+		os.Exit(1)
+	}
+	pub, err := publisher.New(ch, cfg.Exchange, cfg.RoutingKey)
+	if err != nil {
+		slog.Error("publisher init failed", "error", err)
+		os.Exit(1)
+	}
+
+	ctx, stop := signal.NotifyContext(context.Background(), syscall.SIGINT, syscall.SIGTERM)
+	defer stop()
+
+	slog.Info("watcher started",
+		"storage_root", cfg.StorageRoot,
+		"poll_interval", cfg.PollInterval.String(),
+		"exchange", cfg.Exchange,
+		"routing_key", cfg.RoutingKey,
+	)
+
+	ticker := time.NewTicker(cfg.PollInterval)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			slog.Info("watcher stopping")
+			return
+		case <-ticker.C:
+			claimed, err := sc.ScanOnce()
+			if err != nil {
+				slog.Warn("scan failed", "error", err)
+				continue
+			}
+			for _, cf := range claimed {
+				task := publisher.AudioTask{
+					TaskID:   cf.TaskID,
+					FilePath: cf.FilePath,
+					Filename: cf.Filename,
+					Size:     cf.Size,
+				}
+				pubCtx, cancel := context.WithTimeout(ctx, 30*time.Second)
+				err := pub.Publish(pubCtx, task)
+				cancel()
+				if err != nil {
+					slog.Warn("publish failed, rolling back", "task_id", cf.TaskID, "error", err)
+					if rbErr := sc.RollbackToIncoming(cf.FilePath, cf.Filename); rbErr != nil {
+						slog.Error("rollback failed, moving to failed", "task_id", cf.TaskID, "error", rbErr)
+						_ = sc.MoveToFailed(cf.FilePath, cf.Filename)
+					}
+					continue
+				}
+				slog.Info("task published", "task_id", cf.TaskID, "filename", cf.Filename)
+			}
+		}
+	}
+}
+
+func mustRabbit(url string) *amqp.Channel {
+	var conn *amqp.Connection
+	var err error
+	for i := 0; i < 30; i++ {
+		conn, err = amqp.Dial(url)
+		if err == nil {
+			break
+		}
+		slog.Info("waiting for rabbit", "attempt", i+1, "error", err)
+		time.Sleep(2 * time.Second)
+	}
+	if err != nil {
+		slog.Error("rabbit unreachable", "error", err)
+		os.Exit(1)
+	}
+	ch, err := conn.Channel()
+	if err != nil {
+		slog.Error("rabbit channel failed", "error", err)
+		os.Exit(1)
+	}
+	return ch
+}
--- a/watcher/go.mod
+++ b/watcher/go.mod
@@ -0,0 +1,8 @@
+module github.com/yourorg/watcher
+
+go 1.22
+
+require (
+	github.com/oklog/ulid/v2 v2.1.0
+	github.com/rabbitmq/amqp091-go v1.9.0
+)
--- a/watcher/go.sum
+++ b/watcher/go.sum
@@ -0,0 +1,21 @@
+github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/kr/pretty v0.1.0/go.mod h1:dAy3ld7l9f0ibDNOQOHHMYYIIbhfbHSm3C4ZsoJORNo=
+github.com/kr/pty v1.1.1/go.mod h1:pFQYn66WHrOpPYNljwOMqo10TkYh1fy3cYio2l3bCsQ=
+github.com/kr/text v0.1.0/go.mod h1:4Jbv+DJW3UT/LiOwJeYQe1efqtUx/iVham/4vfdArNI=
+github.com/oklog/ulid/v2 v2.1.0 h1:+9lhoxAP56we25tyYETBBY1YLA2SaoLvUFgrP2miPJU=
+github.com/oklog/ulid/v2 v2.1.0/go.mod h1:rcEKHmBBKfef9DhnvX7y1HZBYxjXb0cP5ExxNsTT1QQ=
+github.com/pborman/getopt v0.0.0-20170112200414-7148bc3a4c30/go.mod h1:85jBQOZwpVEaDAr341tbn15RS4fCAsIst0qp7i8ex1o=
+github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
+github.com/rabbitmq/amqp091-go v1.9.0 h1:qrQtyzB4H8BQgEuJwhmVQqVHB9O4+MNDJCCAcpc3Aoo=
+github.com/rabbitmq/amqp091-go v1.9.0/go.mod h1:+jPrT9iY2eLjRaMSRHUhc3z14E/l85kv/f+6luSD3pc=
+github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
+github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=
+github.com/stretchr/testify v1.7.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
+github.com/stretchr/testify v1.8.0/go.mod h1:yNjHg4UonilssWZ8iaSj1OCr/vHnekPRkoO+kdMU+MU=
+go.uber.org/goleak v1.2.1 h1:NBol2c7O1ZokfZ0LEU9K6Whx/KnwvepVetCUhtKja4A=
+go.uber.org/goleak v1.2.1/go.mod h1:qlT2yGI9QafXHhZZLxlSuNsMw3FFLxBr+tBRlmO1xH4=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/check.v1 v1.0.0-20180628173108-788fd7840127/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
+gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
--- a/watcher/internal/config/config.go
+++ b/watcher/internal/config/config.go
@@ -0,0 +1,60 @@
+package config
+
+import (
+	"os"
+	"strconv"
+	"time"
+)
+
+type Config struct {
+	StorageRoot   string
+	IncomingDir   string
+	ProcessingDir string
+	FailedDir     string
+	PollInterval  time.Duration
+	StableWindow  time.Duration
+	StableChecks  int
+	RabbitURL     string
+	Exchange      string
+	RoutingKey    string
+}
+
+func Load() Config {
+	return Config{
+		StorageRoot:   getEnv("STORAGE_ROOT", "/data/storage"),
+		IncomingDir:   getEnv("INCOMING_DIR", "incoming"),
+		ProcessingDir: getEnv("PROCESSING_DIR", "processing"),
+		FailedDir:     getEnv("FAILED_DIR", "failed"),
+		PollInterval:  getDuration("POLL_INTERVAL", 5*time.Second),
+		StableWindow:  getDuration("STABLE_WINDOW", 2*time.Second),
+		StableChecks:  getInt("STABLE_CHECKS", 3),
+		RabbitURL:     getEnv("RABBITMQ_URL", "amqp://guest:guest@localhost:5672/"),
+		Exchange:      getEnv("RABBITMQ_EXCHANGE", "audio_pipeline"),
+		RoutingKey:    getEnv("RABBITMQ_ROUTING_KEY", "audio.new"),
+	}
+}
+
+func getEnv(key, def string) string {
+	if v := os.Getenv(key); v != "" {
+		return v
+	}
+	return def
+}
+
+func getInt(key string, def int) int {
+	if v := os.Getenv(key); v != "" {
+		if i, err := strconv.Atoi(v); err == nil {
+			return i
+		}
+	}
+	return def
+}
+
+func getDuration(key string, def time.Duration) time.Duration {
+	if v := os.Getenv(key); v != "" {
+		if d, err := time.ParseDuration(v); err == nil {
+			return d
+		}
+	}
+	return def
+}
--- a/watcher/internal/publisher/publisher.go
+++ b/watcher/internal/publisher/publisher.go
@@ -0,0 +1,58 @@
+package publisher
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"time"
+
+	amqp "github.com/rabbitmq/amqp091-go"
+)
+
+type AudioTask struct {
+	TaskID    string `json:"task_id"`
+	FilePath  string `json:"file_path"`
+	Filename  string `json:"filename"`
+	Size      int64  `json:"size"`
+	CreatedAt int64  `json:"created_at"`
+}
+
+type Publisher struct {
+	ch         *amqp.Channel
+	exchange   string
+	routingKey string
+}
+
+func New(ch *amqp.Channel, exchange, routingKey string) (*Publisher, error) {
+	if err := ch.Confirm(false); err != nil {
+		return nil, fmt.Errorf("confirm mode: %w", err)
+	}
+	return &Publisher{ch: ch, exchange: exchange, routingKey: routingKey}, nil
+}
+
+func (p *Publisher) Publish(ctx context.Context, task AudioTask) error {
+	if task.CreatedAt == 0 {
+		task.CreatedAt = time.Now().Unix()
+	}
+	body, err := json.Marshal(task)
+	if err != nil {
+		return err
+	}
+	confirms := p.ch.NotifyPublish(make(chan amqp.Confirmation, 1))
+	if err := p.ch.PublishWithContext(ctx, p.exchange, p.routingKey, false, false, amqp.Publishing{
+		ContentType:  "application/json",
+		Body:         body,
+		DeliveryMode: amqp.Persistent,
+	}); err != nil {
+		return err
+	}
+	select {
+	case confirm := <-confirms:
+		if !confirm.Ack {
+			return fmt.Errorf("publish not confirmed")
+		}
+		return nil
+	case <-ctx.Done():
+		return ctx.Err()
+	}
+}
--- a/watcher/internal/scanner/scanner.go
+++ b/watcher/internal/scanner/scanner.go
@@ -0,0 +1,144 @@
+package scanner
+
+import (
+	"fmt"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+
+	"github.com/oklog/ulid/v2"
+)
+
+var allowedExts = map[string]bool{
+	".mp3": true, ".wav": true, ".m4a": true,
+	".ogg": true, ".flac": true, ".webm": true,
+}
+
+type ClaimedFile struct {
+	TaskID   string
+	FilePath string
+	Filename string
+	Size     int64
+}
+
+type Config struct {
+	StorageRoot   string
+	IncomingDir   string
+	ProcessingDir string
+	FailedDir     string
+	StableWindow  time.Duration
+	StableChecks  int
+}
+
+type Scanner struct {
+	cfg Config
+}
+
+func New(cfg Config) *Scanner {
+	return &Scanner{cfg: cfg}
+}
+
+func (s *Scanner) EnsureDirs() error {
+	for _, dir := range []string{s.cfg.IncomingDir, s.cfg.ProcessingDir, s.cfg.FailedDir} {
+		if err := os.MkdirAll(filepath.Join(s.cfg.StorageRoot, dir), 0o755); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (s *Scanner) ScanOnce() ([]ClaimedFile, error) {
+	incoming := filepath.Join(s.cfg.StorageRoot, s.cfg.IncomingDir)
+	entries, err := os.ReadDir(incoming)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return nil, nil
+		}
+		return nil, err
+	}
+
+	var claimed []ClaimedFile
+	for _, e := range entries {
+		if e.IsDir() {
+			continue
+		}
+		name := e.Name()
+		if strings.HasPrefix(name, ".") || strings.HasSuffix(strings.ToLower(name), ".tmp") {
+			continue
+		}
+		ext := strings.ToLower(filepath.Ext(name))
+		if !allowedExts[ext] {
+			continue
+		}
+		src := filepath.Join(incoming, name)
+		if !s.isStable(src) {
+			continue
+		}
+		cf, err := s.claim(src, name, ext)
+		if err != nil {
+			slog.Warn("claim failed", "file", name, "error", err)
+			continue
+		}
+		claimed = append(claimed, cf)
+	}
+	return claimed, nil
+}
+
+func (s *Scanner) isStable(path string) bool {
+	var lastSize int64 = -1
+	for i := 0; i < s.cfg.StableChecks; i++ {
+		info, err := os.Stat(path)
+		if err != nil {
+			return false
+		}
+		size := info.Size()
+		if lastSize >= 0 && size != lastSize {
+			return false
+		}
+		lastSize = size
+		if i < s.cfg.StableChecks-1 {
+			time.Sleep(s.cfg.StableWindow)
+		}
+	}
+	return true
+}
+
+func (s *Scanner) claim(src, originalName, ext string) (ClaimedFile, error) {
+	info, err := os.Stat(src)
+	if err != nil {
+		return ClaimedFile{}, err
+	}
+	taskID := ulid.Make().String()
+	processing := filepath.Join(s.cfg.StorageRoot, s.cfg.ProcessingDir)
+	dst := filepath.Join(processing, taskID+ext)
+	if err := os.Rename(src, dst); err != nil {
+		return ClaimedFile{}, fmt.Errorf("rename: %w", err)
+	}
+	slog.Info("claimed file", "task_id", taskID, "filename", originalName, "path", dst, "size", info.Size())
+	return ClaimedFile{
+		TaskID:   taskID,
+		FilePath: dst,
+		Filename: originalName,
+		Size:     info.Size(),
+	}, nil
+}
+
+func (s *Scanner) RollbackToIncoming(filePath, originalName string) error {
+	incoming := filepath.Join(s.cfg.StorageRoot, s.cfg.IncomingDir)
+	dst := filepath.Join(incoming, originalName)
+	if err := os.Rename(filePath, dst); err != nil {
+		return s.MoveToFailed(filePath, originalName)
+	}
+	return nil
+}
+
+func (s *Scanner) MoveToFailed(filePath, originalName string) error {
+	failed := filepath.Join(s.cfg.StorageRoot, s.cfg.FailedDir)
+	if err := os.MkdirAll(failed, 0o755); err != nil {
+		return err
+	}
+	dst := filepath.Join(failed, originalName)
+	return os.Rename(filePath, dst)
+}
--- a/workers/analyse/Dockerfile
+++ b/workers/analyse/Dockerfile
@@ -0,0 +1,12 @@
+FROM golang:1.22-alpine AS build
+WORKDIR /src
+COPY go.mod go.sum* ./
+RUN go mod download
+COPY . .
+RUN CGO_ENABLED=0 GOOS=linux go build -ldflags="-s -w" -o /analyse ./cmd/analyse
+
+FROM alpine:3.20
+RUN apk add --no-cache ca-certificates
+WORKDIR /app
+COPY --from=build /analyse /app/analyse
+ENTRYPOINT ["/app/analyse"]
--- a/workers/analyse/cmd/analyse/main.go
+++ b/workers/analyse/cmd/analyse/main.go
@@ -0,0 +1,657 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"database/sql"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"net"
+	"net/http"
+	"os"
+	"strings"
+	"time"
+	"unicode/utf8"
+
+	"github.com/joho/godotenv"
+	_ "github.com/jackc/pgx/v5/stdlib"
+	amqp "github.com/rabbitmq/amqp091-go"
+)
+
+func init() {
+	slog.SetDefault(slog.New(slog.NewJSONHandler(os.Stdout, &slog.HandlerOptions{Level: slog.LevelInfo})))
+}
+
+// ── входящее сообщение из очереди analyse (TranscriptionResult от transcribe) ──
+
+type WorkerMessage struct {
+	TaskID        string    `json:"task_id"`
+	Filename      string    `json:"filename"`
+	FilePath      string    `json:"file_path"`
+	Transcription string    `json:"transcription"`
+	Language      string    `json:"language"`
+	Segments      []Segment `json:"segments,omitempty"`
+	Prompts       []Prompt  `json:"prompts"`
+	TranscribedAt int64     `json:"transcribed_at"`
+}
+
+type Segment struct {
+	Start float64 `json:"start"`
+	End   float64 `json:"end"`
+	Text  string  `json:"text"`
+}
+
+type Prompt struct {
+	ID        int    `json:"id"`
+	IDSection int    `json:"id_section"`
+	Name      string `json:"name"`
+	Prompt    string `json:"prompt"`
+	DtCreate  string `json:"dt_create"`
+}
+
+// AnalysisResult — ключ = name промпта, значение = полный JSON-ответ LLM.
+type AnalysisResult map[string]any
+
+// ── LLM request/response ──
+
+type chatMessage struct {
+	Role    string `json:"role"`
+	Content string `json:"content"`
+}
+type chatRequest struct {
+	Model          string  `json:"model"`
+	Temperature    float64 `json:"temperature"`
+	ResponseFormat struct {
+		Type string `json:"type"`
+	} `json:"response_format"`
+	Messages []chatMessage `json:"messages"`
+}
+type tokenUsage struct {
+	PromptTokens     int `json:"prompt_tokens"`
+	CompletionTokens int `json:"completion_tokens"`
+	TotalTokens      int `json:"total_tokens"`
+}
+type chatResponse struct {
+	Choices []struct {
+		Message struct {
+			Content string `json:"content"`
+		} `json:"message"`
+	} `json:"choices"`
+	Usage *tokenUsage `json:"usage"`
+}
+
+type llmCallResult struct {
+	Content       string
+	RequestBytes  int
+	ResponseBytes int
+	Usage         *tokenUsage
+	Duration      time.Duration
+}
+
+type analysisStats struct {
+	LLMCalls      int
+	TotalTokens   int
+	PromptTokens  int
+	OutputTokens  int
+}
+
+// ===================== LLM =====================
+
+var llmHTTPClient = newLLMHTTPClient(150 * time.Second)
+
+func newLLMHTTPClient(totalTimeout time.Duration) *http.Client {
+	return &http.Client{
+		Timeout: totalTimeout,
+		Transport: &http.Transport{
+			Proxy: http.ProxyFromEnvironment,
+			DialContext: (&net.Dialer{
+				Timeout:   30 * time.Second,
+				KeepAlive: 30 * time.Second,
+			}).DialContext,
+			TLSHandshakeTimeout:   60 * time.Second,
+			ResponseHeaderTimeout: 90 * time.Second,
+			ExpectContinueTimeout: 5 * time.Second,
+			IdleConnTimeout:       90 * time.Second,
+		},
+	}
+}
+
+func callLLM(ctx context.Context, apiURL, model, prompt string) (*llmCallResult, error) {
+	const systemPrompt = "Ты — строгий классификатор звонков. Отвечай только JSON, без пояснений."
+
+	reqBody := chatRequest{
+		Model:       model,
+		Temperature: 0.1,
+		Messages: []chatMessage{
+			{Role: "system", Content: systemPrompt},
+			{Role: "user", Content: prompt},
+		},
+	}
+	reqBody.ResponseFormat.Type = "json_object"
+
+	jsonData, err := json.Marshal(reqBody)
+	if err != nil {
+		return nil, err
+	}
+
+	req, err := http.NewRequestWithContext(ctx, "POST", apiURL, bytes.NewBuffer(jsonData))
+	if err != nil {
+		return nil, err
+	}
+	req.Header.Set("Authorization", "Bearer "+os.Getenv("YANDEX_API_KEY"))
+	req.Header.Set("Content-Type", "application/json")
+
+	start := time.Now()
+	resp, err := llmHTTPClient.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close()
+
+	body, _ := io.ReadAll(resp.Body)
+	duration := time.Since(start)
+
+	if resp.StatusCode != http.StatusOK {
+		return &llmCallResult{
+			RequestBytes:  len(jsonData),
+			ResponseBytes: len(body),
+			Duration:      duration,
+		}, fmt.Errorf("status %d: %s", resp.StatusCode, truncate(string(body), 500))
+	}
+
+	var result chatResponse
+	if err := json.Unmarshal(body, &result); err != nil {
+		return &llmCallResult{
+			RequestBytes:  len(jsonData),
+			ResponseBytes: len(body),
+			Duration:      duration,
+		}, err
+	}
+	if len(result.Choices) == 0 {
+		return &llmCallResult{
+			RequestBytes:  len(jsonData),
+			ResponseBytes: len(body),
+			Duration:      duration,
+		}, fmt.Errorf("empty response")
+	}
+
+	return &llmCallResult{
+		Content:       result.Choices[0].Message.Content,
+		RequestBytes:  len(jsonData),
+		ResponseBytes: len(body),
+		Usage:         result.Usage,
+		Duration:      duration,
+	}, nil
+}
+
+func checkYandexAPI(ctx context.Context, apiURL, model string) error {
+	slog.Info("yandex api check started", "worker", "analyse", "url", apiURL, "model", model)
+
+	res, err := callLLM(ctx, apiURL, model, `Ответь только JSON: {"ok":true}`)
+	if err != nil {
+		return err
+	}
+
+	attrs := []any{
+		"worker", "analyse",
+		"duration_ms", res.Duration.Milliseconds(),
+		"response_chars", utf8.RuneCountInString(res.Content),
+	}
+	if res.Usage != nil {
+		attrs = append(attrs,
+			"prompt_tokens", res.Usage.PromptTokens,
+			"completion_tokens", res.Usage.CompletionTokens,
+			"total_tokens", res.Usage.TotalTokens,
+		)
+	}
+	slog.Info("yandex api check ok", attrs...)
+	return nil
+}
+
+func logLLMCall(taskID, model, promptName string, promptIndex, promptTotal, attempt, inputChars int, res *llmCallResult, err error) {
+	attrs := []any{
+		"worker", "analyse",
+		"task_id", taskID,
+		"model", model,
+		"call_type", "analyse_prompt",
+		"prompt_name", promptName,
+		"prompt_index", promptIndex,
+		"prompt_total", promptTotal,
+		"attempt", attempt,
+		"input_chars", inputChars,
+	}
+	if res != nil {
+		attrs = append(attrs,
+			"duration_ms", res.Duration.Milliseconds(),
+			"request_bytes", res.RequestBytes,
+			"response_bytes", res.ResponseBytes,
+			"response_chars", utf8.RuneCountInString(res.Content),
+		)
+		if res.Usage != nil {
+			attrs = append(attrs,
+				"prompt_tokens", res.Usage.PromptTokens,
+				"completion_tokens", res.Usage.CompletionTokens,
+				"total_tokens", res.Usage.TotalTokens,
+			)
+		}
+	}
+	if err != nil {
+		slog.Warn("llm call failed", append(attrs, "error", err)...)
+		return
+	}
+	slog.Info("llm call ok", attrs...)
+}
+
+func accumulateUsage(stats *analysisStats, res *llmCallResult) {
+	stats.LLMCalls++
+	if res != nil && res.Usage != nil {
+		stats.TotalTokens += res.Usage.TotalTokens
+		stats.PromptTokens += res.Usage.PromptTokens
+		stats.OutputTokens += res.Usage.CompletionTokens
+	}
+}
+
+func buildPromptQuery(transcription string, p Prompt) string {
+	var b strings.Builder
+	b.WriteString(p.Prompt)
+	b.WriteString("\n\n=== ТРАНСКРИПЦИЯ ===\n\"\"\"\n")
+	b.WriteString(transcription)
+	b.WriteString("\n\"\"\"")
+	return b.String()
+}
+
+func analysePrompt(ctx context.Context, apiURL, model, transcription string, p Prompt, index, total int, taskID string, stats *analysisStats) (any, error) {
+	query := buildPromptQuery(transcription, p)
+	inputChars := utf8.RuneCountInString(query)
+
+	res, err := callLLM(ctx, apiURL, model, query)
+	logLLMCall(taskID, model, p.Name, index, total, 1, inputChars, res, err)
+	accumulateUsage(stats, res)
+	if err != nil {
+		return nil, err
+	}
+
+	var parsed any
+	if err := json.Unmarshal([]byte(res.Content), &parsed); err != nil {
+		return nil, fmt.Errorf("parse: %w, resp: %s", err, truncate(res.Content, 300))
+	}
+	return parsed, nil
+}
+
+func runAnalysis(ctx context.Context, apiURL, model, taskID, transcription string, prompts []Prompt) (AnalysisResult, analysisStats, error) {
+	stats := analysisStats{}
+	result := make(AnalysisResult, len(prompts))
+
+	valid := make([]Prompt, 0, len(prompts))
+	for _, p := range prompts {
+		if p.Name != "" {
+			valid = append(valid, p)
+		}
+	}
+	total := len(valid)
+
+	for i, p := range valid {
+		value, err := analysePrompt(ctx, apiURL, model, transcription, p, i+1, total, taskID, &stats)
+		if err != nil {
+			return nil, stats, fmt.Errorf("%s: %w", p.Name, err)
+		}
+		result[p.Name] = value
+	}
+	return result, stats, nil
+}
+
+// ===================== DB =====================
+
+func saveAnalysis(ctx context.Context, db *sql.DB, task WorkerMessage, analysis []byte) (complete bool, err error) {
+	metadata, _ := json.Marshal(map[string]any{
+		"file_path":      task.FilePath,
+		"language":       task.Language,
+		"segments":       task.Segments,
+		"prompts":        task.Prompts,
+		"transcribed_at": task.TranscribedAt,
+	})
+
+	_, err = db.ExecContext(ctx,
+		`INSERT INTO results (task_id) VALUES ($1) ON CONFLICT (task_id) DO NOTHING`, task.TaskID)
+	if err != nil {
+		return false, fmt.Errorf("ensure row: %w", err)
+	}
+
+	err = db.QueryRowContext(ctx, `
+		UPDATE results
+		SET analysis = $2::jsonb,
+		    filename = COALESCE(NULLIF($3, ''), filename),
+		    transcription = COALESCE(NULLIF($4, ''), transcription),
+		    metadata = COALESCE($5::jsonb, metadata),
+		    updated_at = now(),
+		    status = CASE WHEN tagging IS NOT NULL THEN 'done' ELSE status END
+		WHERE task_id = $1
+		RETURNING (analysis IS NOT NULL AND tagging IS NOT NULL)
+	`, task.TaskID, string(analysis), task.Filename, task.Transcription, string(metadata)).Scan(&complete)
+	if err != nil {
+		return false, fmt.Errorf("update analysis: %w", err)
+	}
+	return complete, nil
+}
+
+// ===================== MAIN =====================
+
+func loadDotenv() {
+	path := os.Getenv("DOTENV_PATH")
+	if path == "" {
+		return
+	}
+	if err := godotenv.Overload(path); err != nil {
+		slog.Warn("dotenv load failed", "path", path, "error", err)
+		return
+	}
+	slog.Info("dotenv loaded", "path", path)
+}
+
+func main() {
+	loadDotenv()
+
+	amqpURL := getEnv("RABBITMQ_URL", "amqp://guest:guest@localhost:5672/")
+	dbURL := getEnv("DATABASE_URL", "")
+	token := os.Getenv("YANDEX_API_KEY")
+	model := os.Getenv("YANDEX_MODEL")
+	apiURL := getEnv("YANDEX_API_URL", "https://ai.api.cloud.yandex.net/v1/chat/completions")
+	inputQueue := getEnv("ANALYSE_QUEUE", "analyse")
+	finalQueue := getEnv("FINAL_QUEUE", "final")
+
+	if token == "" {
+		slog.Error("YANDEX_API_KEY is required")
+		os.Exit(1)
+	}
+	if model == "" {
+		slog.Error("YANDEX_MODEL is required")
+		os.Exit(1)
+	}
+	if dbURL == "" {
+		slog.Error("DATABASE_URL is required")
+		os.Exit(1)
+	}
+	slog.Info("config loaded", "worker", "analyse",
+		"yandex_token", tokenFingerprint(token), "model", model, "api_url", apiURL)
+
+	db := mustDB(dbURL)
+	defer db.Close()
+
+	checkCtx, checkCancel := context.WithTimeout(context.Background(), 90*time.Second)
+	if err := checkYandexAPI(checkCtx, apiURL, model); err != nil {
+		checkCancel()
+		slog.Error("yandex api check failed — worker will not start", "worker", "analyse", "error", err)
+		os.Exit(1)
+	}
+	checkCancel()
+
+	ch := mustRabbit(amqpURL)
+
+	if _, err := ch.QueueDeclare(inputQueue, true, false, false, false, nil); err != nil {
+		slog.Error("declare queue failed", "queue", inputQueue, "error", err)
+		os.Exit(1)
+	}
+	if _, err := ch.QueueDeclare(finalQueue, true, false, false, false, nil); err != nil {
+		slog.Error("declare queue failed", "queue", finalQueue, "error", err)
+		os.Exit(1)
+	}
+	ch.Qos(1, 0, false)
+
+	msgs, err := ch.Consume(inputQueue, "", false, false, false, false, nil)
+	if err != nil {
+		slog.Error("consume failed", "error", err)
+		os.Exit(1)
+	}
+	slog.Info("worker started", "worker", "analyse", "queue", inputQueue, "model", model)
+
+	for d := range msgs {
+		taskStart := time.Now()
+		var task WorkerMessage
+		if err := json.Unmarshal(d.Body, &task); err != nil {
+			slog.Warn("bad message", "worker", "analyse", "delivery_tag", d.DeliveryTag,
+				"body_bytes", len(d.Body), "error", err)
+			d.Nack(false, false)
+			continue
+		}
+
+		promptNames := make([]string, 0, len(task.Prompts))
+		promptTextChars := 0
+		for _, p := range task.Prompts {
+			if p.Name != "" {
+				promptNames = append(promptNames, p.Name)
+				promptTextChars += utf8.RuneCountInString(p.Prompt)
+			}
+		}
+		transcriptionChars := utf8.RuneCountInString(task.Transcription)
+
+		slog.Info("message received", "worker", "analyse",
+			"task_id", task.TaskID,
+			"filename", task.Filename,
+			"delivery_tag", d.DeliveryTag,
+			"redelivered", d.Redelivered,
+			"body_bytes", len(d.Body),
+			"transcription_chars", transcriptionChars,
+			"segments", len(task.Segments),
+			"prompts", len(promptNames),
+			"prompt_names", promptNames,
+			"prompt_text_chars", promptTextChars,
+			"llm_calls_expected", len(promptNames),
+		)
+		if d.Redelivered {
+			slog.Warn("redelivered message skipped — no llm call",
+				"worker", "analyse", "task_id", task.TaskID,
+				"delivery_tag", d.DeliveryTag, "prompts", len(promptNames))
+			d.Nack(false, false)
+			continue
+		}
+
+		ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
+
+		result, stats, err := runAnalysis(ctx, apiURL, model, task.TaskID, task.Transcription, task.Prompts)
+		if err != nil {
+			cancel()
+			slog.Warn("task failed, discarded",
+				"worker", "analyse", "task_id", task.TaskID,
+				"llm_calls_done", stats.LLMCalls,
+				"total_tokens_so_far", stats.TotalTokens,
+				"error", err)
+			d.Nack(false, false)
+			continue
+		}
+
+		analysisJSON, _ := json.Marshal(result)
+		complete, err := saveAnalysis(ctx, db, task, analysisJSON)
+		if err != nil {
+			cancel()
+			slog.Warn("db save failed, discarded",
+				"worker", "analyse", "task_id", task.TaskID, "error", err)
+			d.Nack(false, false)
+			continue
+		}
+
+		taskAttrs := []any{
+			"worker", "analyse",
+			"task_id", task.TaskID,
+			"llm_calls", stats.LLMCalls,
+			"total_tokens", stats.TotalTokens,
+			"prompt_tokens", stats.PromptTokens,
+			"completion_tokens", stats.OutputTokens,
+			"duration_ms", time.Since(taskStart).Milliseconds(),
+		}
+
+		if complete {
+			notifyFinal(ctx, ch, db, finalQueue, task.TaskID, "analyse")
+			slog.Info("task complete", append(taskAttrs, "was_last", "analyse")...)
+		} else {
+			slog.Info("task partial", append(taskAttrs, "waiting_for", "tagging")...)
+		}
+		cancel()
+
+		d.Ack(false)
+	}
+}
+
+func truncate(s string, max int) string {
+	if len(s) <= max {
+		return s
+	}
+	return s[:max] + "..."
+}
+
+func loadFinalPayload(ctx context.Context, db *sql.DB, taskID string) ([]byte, error) {
+	var (
+		filename, transcription, status sql.NullString
+		analysis, tagging, metadata     []byte
+		createdAt, updatedAt            time.Time
+	)
+	err := db.QueryRowContext(ctx, `
+		SELECT filename, transcription, analysis, tagging, metadata, status, created_at, updated_at
+		FROM results WHERE task_id = $1
+	`, taskID).Scan(&filename, &transcription, &analysis, &tagging, &metadata, &status, &createdAt, &updatedAt)
+	if err != nil {
+		return nil, fmt.Errorf("load result: %w", err)
+	}
+
+	msg := map[string]any{
+		"task_id":    taskID,
+		"status":     status.String,
+		"created_at": createdAt,
+		"updated_at": updatedAt,
+	}
+	if filename.Valid {
+		msg["filename"] = filename.String
+	}
+	if transcription.Valid {
+		msg["transcription"] = transcription.String
+	}
+	if len(analysis) > 0 {
+		var v any
+		if err := json.Unmarshal(analysis, &v); err == nil {
+			msg["analysis"] = v
+		}
+	}
+	if len(tagging) > 0 {
+		var v any
+		if err := json.Unmarshal(tagging, &v); err == nil {
+			msg["tagging"] = v
+		}
+	}
+	if len(metadata) > 0 {
+		var meta map[string]any
+		if err := json.Unmarshal(metadata, &meta); err == nil {
+			for k, v := range meta {
+				msg[k] = v
+			}
+		}
+	}
+	return json.Marshal(msg)
+}
+
+func notifyFinal(ctx context.Context, ch *amqp.Channel, db *sql.DB, queue, taskID, worker string) {
+	body, err := loadFinalPayload(ctx, db, taskID)
+	if err != nil {
+		slog.Warn("load final payload failed", "worker", worker, "task_id", taskID, "error", err)
+		return
+	}
+	if err := ch.PublishWithContext(ctx, "", queue, false, false,
+		amqp.Publishing{
+			ContentType:  "application/json",
+			Body:         body,
+			DeliveryMode: amqp.Persistent,
+		}); err != nil {
+		slog.Warn("publish final failed", "worker", worker, "task_id", taskID, "error", err)
+		return
+	}
+	slog.Info("published final", "worker", worker, "task_id", taskID, "queue", queue, "body_bytes", len(body))
+	deleteProcessingFile(extractFilePath(body), taskID, worker)
+}
+
+func extractFilePath(body []byte) string {
+	var msg map[string]any
+	if err := json.Unmarshal(body, &msg); err != nil {
+		return ""
+	}
+	fp, _ := msg["file_path"].(string)
+	return fp
+}
+
+func deleteProcessingFile(filePath, taskID, worker string) {
+	if filePath == "" {
+		slog.Warn("processing file not deleted: no file_path", "worker", worker, "task_id", taskID)
+		return
+	}
+	if !strings.Contains(filePath, "/processing/") {
+		slog.Warn("processing file not deleted: path outside processing", "worker", worker, "task_id", taskID, "path", filePath)
+		return
+	}
+	if err := os.Remove(filePath); err != nil {
+		if os.IsNotExist(err) {
+			slog.Info("processing file already removed", "worker", worker, "task_id", taskID, "path", filePath)
+			return
+		}
+		slog.Warn("processing file delete failed", "worker", worker, "task_id", taskID, "path", filePath, "error", err)
+		return
+	}
+	slog.Info("processing file deleted", "worker", worker, "task_id", taskID, "path", filePath)
+}
+
+func getEnv(k, d string) string {
+	if v := os.Getenv(k); v != "" {
+		return v
+	}
+	return d
+}
+
+func tokenFingerprint(token string) string {
+	if len(token) <= 12 {
+		return "***"
+	}
+	return token[:8] + "..." + token[len(token)-4:]
+}
+
+func mustDB(url string) *sql.DB {
+	db, err := sql.Open("pgx", url)
+	if err != nil {
+		slog.Error("db open failed", "error", err)
+		os.Exit(1)
+	}
+	db.SetMaxOpenConns(5)
+	time.Sleep(2 * time.Second) // дать Docker DNS зарегистрировать postgres
+	for i := 0; i < 60; i++ {
+		if err = db.Ping(); err == nil {
+			return db
+		}
+		if i < 5 || (i+1)%10 == 0 {
+			slog.Info("waiting for db", "attempt", i+1, "error", err)
+		}
+		time.Sleep(3 * time.Second)
+	}
+	slog.Error("db unreachable", "error", err)
+	os.Exit(1)
+	return nil
+}
+
+func mustRabbit(url string) *amqp.Channel {
+	var conn *amqp.Connection
+	var err error
+	for i := 0; i < 30; i++ {
+		conn, err = amqp.Dial(url)
+		if err == nil {
+			break
+		}
+		slog.Info("waiting for rabbit", "attempt", i+1, "error", err)
+		time.Sleep(2 * time.Second)
+	}
+	if err != nil {
+		slog.Error("rabbit unreachable", "error", err)
+		os.Exit(1)
+	}
+	ch, err := conn.Channel()
+	if err != nil {
+		slog.Error("rabbit channel failed", "error", err)
+		os.Exit(1)
+	}
+	return ch
+}
--- a/workers/analyse/go.mod
+++ b/workers/analyse/go.mod
@@ -0,0 +1,18 @@
+module github.com/yourorg/analyse
+
+go 1.22
+
+require (
+	github.com/jackc/pgx/v5 v5.5.5
+	github.com/joho/godotenv v1.5.1
+	github.com/rabbitmq/amqp091-go v1.9.0
+)
+
+require (
+	github.com/jackc/pgpassfile v1.0.0 // indirect
+	github.com/jackc/pgservicefile v0.0.0-20221227161230-091c0ba34f0a // indirect
+	github.com/jackc/puddle/v2 v2.2.1 // indirect
+	golang.org/x/crypto v0.17.0 // indirect
+	golang.org/x/sync v0.1.0 // indirect
+	golang.org/x/text v0.14.0 // indirect
+)
--- a/workers/analyse/go.sum
+++ b/workers/analyse/go.sum
@@ -0,0 +1,41 @@
+github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
+github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/jackc/pgpassfile v1.0.0 h1:/6Hmqy13Ss2zCq62VdNG8tM1wchn8zjSGOBJ6icpsIM=
+github.com/jackc/pgpassfile v1.0.0/go.mod h1:CEx0iS5ambNFdcRtxPj5JhEz+xB6uRky5eyVu/W2HEg=
+github.com/jackc/pgservicefile v0.0.0-20221227161230-091c0ba34f0a h1:bbPeKD0xmW/Y25WS6cokEszi5g+S0QxI/d45PkRi7Nk=
+github.com/jackc/pgservicefile v0.0.0-20221227161230-091c0ba34f0a/go.mod h1:5TJZWKEWniPve33vlWYSoGYefn3gLQRzjfDlhSJ9ZKM=
+github.com/jackc/pgx/v5 v5.5.5 h1:amBjrZVmksIdNjxGW/IiIMzxMKZFelXbUoPNb+8sjQw=
+github.com/jackc/pgx/v5 v5.5.5/go.mod h1:ez9gk+OAat140fv9ErkZDYFWmXLfV+++K0uAOiwgm1A=
+github.com/jackc/puddle/v2 v2.2.1 h1:RhxXJtFG022u4ibrCSMSiu5aOq1i77R3OHKNJj77OAk=
+github.com/jackc/puddle/v2 v2.2.1/go.mod h1:vriiEXHvEE654aYKXXjOvZM39qJ0q+azkZFrfEOc3H4=
+github.com/joho/godotenv v1.5.1 h1:7eLL/+HRGLY0ldzfGMeQkb7vMd0as4CfYvUVzLqw0N0=
+github.com/joho/godotenv v1.5.1/go.mod h1:f4LDr5Voq0i2e/R5DDNOoa2zzDfwtkZa6DnEwAbqwq4=
+github.com/kr/pretty v0.1.0/go.mod h1:dAy3ld7l9f0ibDNOQOHHMYYIIbhfbHSm3C4ZsoJORNo=
+github.com/kr/pty v1.1.1/go.mod h1:pFQYn66WHrOpPYNljwOMqo10TkYh1fy3cYio2l3bCsQ=
+github.com/kr/text v0.1.0/go.mod h1:4Jbv+DJW3UT/LiOwJeYQe1efqtUx/iVham/4vfdArNI=
+github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
+github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
+github.com/rabbitmq/amqp091-go v1.9.0 h1:qrQtyzB4H8BQgEuJwhmVQqVHB9O4+MNDJCCAcpc3Aoo=
+github.com/rabbitmq/amqp091-go v1.9.0/go.mod h1:+jPrT9iY2eLjRaMSRHUhc3z14E/l85kv/f+6luSD3pc=
+github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
+github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=
+github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=
+github.com/stretchr/testify v1.7.0/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
+github.com/stretchr/testify v1.7.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
+github.com/stretchr/testify v1.8.0/go.mod h1:yNjHg4UonilssWZ8iaSj1OCr/vHnekPRkoO+kdMU+MU=
+github.com/stretchr/testify v1.8.1 h1:w7B6lhMri9wdJUVmEZPGGhZzrYTPvgJArz7wNPgYKsk=
+github.com/stretchr/testify v1.8.1/go.mod h1:w2LPCIKwWwSfY2zedu0+kehJoqGctiVI29o6fzry7u4=
+go.uber.org/goleak v1.2.1 h1:NBol2c7O1ZokfZ0LEU9K6Whx/KnwvepVetCUhtKja4A=
+go.uber.org/goleak v1.2.1/go.mod h1:qlT2yGI9QafXHhZZLxlSuNsMw3FFLxBr+tBRlmO1xH4=
+golang.org/x/crypto v0.17.0 h1:r8bRNjWL3GshPW3gkd+RpvzWrZAwPS49OmTGZ/uhM4k=
+golang.org/x/crypto v0.17.0/go.mod h1:gCAAfMLgwOJRpTjQ2zCCt2OcSfYMTeZVSRtQlPC7Nq4=
+golang.org/x/sync v0.1.0 h1:wsuoTGHzEhffawBOhz5CYhcrV4IdKZbEyZjBMuTp12o=
+golang.org/x/sync v0.1.0/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
+golang.org/x/text v0.14.0 h1:ScX5w1eTa3QqT8oi6+ziP7dTV1S2+ALU0bI+0zXKWiQ=
+golang.org/x/text v0.14.0/go.mod h1:18ZOQIKpY8NJVqYksKHtTdi31H5itFRjB5/qKTNYzSU=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/check.v1 v1.0.0-20180628173108-788fd7840127/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
+gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
+gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
--- a/workers/tagging/Dockerfile
+++ b/workers/tagging/Dockerfile
@@ -0,0 +1,11 @@
+FROM golang:1.22-alpine AS build
+WORKDIR /src
+COPY go.mod go.sum* ./
+RUN go mod download
+COPY . .
+RUN CGO_ENABLED=0 go build -o /tagging ./cmd/tagging
+
+FROM alpine:3.19
+RUN apk add --no-cache ca-certificates
+COPY --from=build /tagging /tagging
+ENTRYPOINT ["/tagging"]
--- a/workers/tagging/cmd/tagging/main.go
+++ b/workers/tagging/cmd/tagging/main.go
@@ -0,0 +1,685 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"database/sql"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"net"
+	"net/http"
+	"os"
+	"strings"
+	"time"
+	"unicode/utf8"
+
+	"github.com/joho/godotenv"
+	_ "github.com/jackc/pgx/v5/stdlib"
+	amqp "github.com/rabbitmq/amqp091-go"
+)
+
+func init() {
+	slog.SetDefault(slog.New(slog.NewJSONHandler(os.Stdout, &slog.HandlerOptions{Level: slog.LevelInfo})))
+}
+
+func apiURL() string {
+	if u := os.Getenv("YANDEX_API_URL"); u != "" {
+		return u
+	}
+	return "https://ai.api.cloud.yandex.net/v1/chat/completions"
+}
+
+// ── входящее сообщение из очереди tagging ──
+type WorkerMessage struct {
+	TaskID        string `json:"task_id"`
+	Filename      string `json:"filename"`
+	Transcription string `json:"transcription"`
+}
+
+// ── результат классификации ──
+type ClassificationResult struct {
+	L1             string `json:"L1"`
+	L2             string `json:"L2"`
+	L3             string `json:"L3"`
+	RiskLevel      string `json:"risk_level"`
+	HasActionItems bool   `json:"has_action_items"`
+	HasDeadline    bool   `json:"has_deadline"`
+}
+
+// ── LLM request/response ──
+type chatMessage struct {
+	Role    string `json:"role"`
+	Content string `json:"content"`
+}
+type chatRequest struct {
+	Model          string  `json:"model"`
+	Temperature    float64 `json:"temperature"`
+	ResponseFormat struct {
+		Type string `json:"type"`
+	} `json:"response_format"`
+	Messages []chatMessage `json:"messages"`
+}
+type tokenUsage struct {
+	PromptTokens     int `json:"prompt_tokens"`
+	CompletionTokens int `json:"completion_tokens"`
+	TotalTokens      int `json:"total_tokens"`
+}
+type chatResponse struct {
+	Choices []struct {
+		Message struct {
+			Content string `json:"content"`
+		} `json:"message"`
+	} `json:"choices"`
+	Usage *tokenUsage `json:"usage"`
+}
+
+type llmCallResult struct {
+	Content       string
+	RequestBytes  int
+	ResponseBytes int
+	Usage         *tokenUsage
+	Duration      time.Duration
+}
+
+// ===================== LLM =====================
+
+var llmHTTPClient = newLLMHTTPClient(90 * time.Second)
+
+func newLLMHTTPClient(totalTimeout time.Duration) *http.Client {
+	return &http.Client{
+		Timeout: totalTimeout,
+		Transport: &http.Transport{
+			Proxy: http.ProxyFromEnvironment,
+			DialContext: (&net.Dialer{
+				Timeout:   30 * time.Second,
+				KeepAlive: 30 * time.Second,
+			}).DialContext,
+			TLSHandshakeTimeout:   60 * time.Second,
+			ResponseHeaderTimeout: 60 * time.Second,
+			ExpectContinueTimeout: 5 * time.Second,
+			IdleConnTimeout:       90 * time.Second,
+		},
+	}
+}
+
+func callLLM(ctx context.Context, model, prompt string) (*llmCallResult, error) {
+	const systemPrompt = "Ты — классификатор диалогов в логистике. Отвечай только JSON, без пояснений."
+
+	reqBody := chatRequest{
+		Model:       model,
+		Temperature: 0.1,
+		Messages: []chatMessage{
+			{Role: "system", Content: systemPrompt},
+			{Role: "user", Content: prompt},
+		},
+	}
+	reqBody.ResponseFormat.Type = "json_object"
+
+	jsonData, err := json.Marshal(reqBody)
+	if err != nil {
+		return nil, err
+	}
+
+	req, err := http.NewRequestWithContext(ctx, "POST", apiURL(), bytes.NewBuffer(jsonData))
+	if err != nil {
+		return nil, err
+	}
+	req.Header.Set("Authorization", "Bearer "+os.Getenv("YANDEX_API_KEY"))
+	req.Header.Set("Content-Type", "application/json")
+
+	start := time.Now()
+	resp, err := llmHTTPClient.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close()
+
+	body, _ := io.ReadAll(resp.Body)
+	duration := time.Since(start)
+
+	if resp.StatusCode != http.StatusOK {
+		return &llmCallResult{
+			RequestBytes:  len(jsonData),
+			ResponseBytes: len(body),
+			Duration:      duration,
+		}, fmt.Errorf("status %d: %s", resp.StatusCode, truncate(string(body), 500))
+	}
+
+	var result chatResponse
+	if err := json.Unmarshal(body, &result); err != nil {
+		return &llmCallResult{
+			RequestBytes:  len(jsonData),
+			ResponseBytes: len(body),
+			Duration:      duration,
+		}, err
+	}
+	if len(result.Choices) == 0 {
+		return &llmCallResult{
+			RequestBytes:  len(jsonData),
+			ResponseBytes: len(body),
+			Duration:      duration,
+		}, fmt.Errorf("empty response")
+	}
+
+	return &llmCallResult{
+		Content:       result.Choices[0].Message.Content,
+		RequestBytes:  len(jsonData),
+		ResponseBytes: len(body),
+		Usage:         result.Usage,
+		Duration:      duration,
+	}, nil
+}
+
+func checkYandexAPI(ctx context.Context, model string) error {
+	slog.Info("yandex api check started", "worker", "tagging", "url", apiURL(), "model", model)
+
+	res, err := callLLM(ctx, model, `Ответь только JSON: {"ok":true}`)
+	if err != nil {
+		return err
+	}
+
+	attrs := []any{
+		"worker", "tagging",
+		"duration_ms", res.Duration.Milliseconds(),
+		"response_chars", utf8.RuneCountInString(res.Content),
+	}
+	if res.Usage != nil {
+		attrs = append(attrs,
+			"prompt_tokens", res.Usage.PromptTokens,
+			"completion_tokens", res.Usage.CompletionTokens,
+			"total_tokens", res.Usage.TotalTokens,
+		)
+	}
+	slog.Info("yandex api check ok", attrs...)
+	return nil
+}
+
+func logLLMCall(worker, taskID, model, callType string, attempt int, inputChars int, res *llmCallResult, err error) {
+	attrs := []any{
+		"worker", worker,
+		"task_id", taskID,
+		"model", model,
+		"call_type", callType,
+		"attempt", attempt,
+		"input_chars", inputChars,
+	}
+	if res != nil {
+		attrs = append(attrs,
+			"duration_ms", res.Duration.Milliseconds(),
+			"request_bytes", res.RequestBytes,
+			"response_bytes", res.ResponseBytes,
+			"response_chars", utf8.RuneCountInString(res.Content),
+		)
+		if res.Usage != nil {
+			attrs = append(attrs,
+				"prompt_tokens", res.Usage.PromptTokens,
+				"completion_tokens", res.Usage.CompletionTokens,
+				"total_tokens", res.Usage.TotalTokens,
+			)
+		}
+	}
+	if err != nil {
+		slog.Warn("llm call failed", append(attrs, "error", err)...)
+		return
+	}
+	slog.Info("llm call ok", attrs...)
+}
+
+func buildPrompt(text string) string {
+	return fmt.Sprintf(`Ты — классификатор диалогов в логистике.
+
+Тебе даётся НЕструктурированный текст диалога (разговор, звонок, переписка).
+Текст может быть неаккуратным, с ошибками, без структуры.
+
+Твоя задача:
+1. Понять смысл диалога
+2. Выделить ключевую цель разговора
+3. Определить наличие проблемы
+4. Классифицировать диалог по правилам ниже
+
+=== ИЕРАРХИЯ КЛАССОВ ===
+
+L1:
+- new_order
+- order_change
+- tracking
+- delivery_coordination
+- problem
+- claim
+- information_request
+- internal_communication
+- other
+
+L2:
+
+Для problem:
+- delivery_issue
+- cargo_issue
+- data_issue
+- communication_issue
+
+Для delivery_coordination:
+- delivery_time
+- unloading_conditions
+- warehouse_rules
+- access
+- scheduling
+
+Для tracking:
+- location_request
+- status_update
+- eta
+
+L3 (опционально):
+- wrong_contact
+- wrong_address
+- missing_info
+- delay
+- lost
+- damage
+- cannot_reach
+- no_response
+
+=== ДОПОЛНИТЕЛЬНЫЕ ПОЛЯ ===
+
+risk_level:
+- none
+- low
+- medium
+- high
+
+has_action_items:
+- true / false
+
+has_deadline:
+- true / false
+
+=== ПРАВИЛА ===
+
+1. Определи основную цель разговора:
+   - заказ → new_order
+   - изменение → order_change
+   - узнать статус → tracking
+   - согласование → delivery_coordination
+   - ошибка / проблема → problem
+
+2. Если есть любая ошибка или сбой → ВСЕГДА L1 = problem
+
+3. Ошибки в email / телефоне / адресе → L2 = data_issue
+
+4. Если обсуждают условия (время, склад, разгрузка) без проблемы → delivery_coordination
+
+5. Если спрашивают "где груз?" → tracking
+
+6. Определи risk_level:
+   - low → проблема не влияет на доставку
+   - medium → возможна задержка
+   - high → срыв сроков / потеря
+
+7. has_action_items = true если:
+   - есть договорённости ("перезвоню", "свяжется", "отправлю")
+
+8. has_deadline = true если:
+   - есть конкретное время ("в 18:00", "через 10 минут", "завтра")
+
+---
+
+=== ФОРМАТ ОТВЕТА ===
+
+Ответ только JSON, без пояснений:
+
+{
+  "L1": "...",
+  "L2": "...",
+  "L3": "...",
+  "risk_level": "...",
+  "has_action_items": true/false,
+  "has_deadline": true/false
+}
+
+---
+
+=== ДИАЛОГ ===
+
+Текст:
+"""
+%s
+"""`, text)
+}
+
+func classify(ctx context.Context, taskID, model, text string) (ClassificationResult, error) {
+	prompt := buildPrompt(text)
+	inputChars := utf8.RuneCountInString(prompt)
+
+	res, err := callLLM(ctx, model, prompt)
+	logLLMCall("tagging", taskID, model, "classify", 1, inputChars, res, err)
+	if err != nil {
+		return ClassificationResult{}, err
+	}
+
+	var result ClassificationResult
+	if err := json.Unmarshal([]byte(res.Content), &result); err != nil {
+		return ClassificationResult{}, fmt.Errorf("parse: %w, resp: %s", err, truncate(res.Content, 300))
+	}
+	return result, nil
+}
+
+// ===================== DB =====================
+
+func saveTagging(ctx context.Context, db *sql.DB, taskID, filename, transcription string, tagging []byte) (complete bool, err error) {
+	_, err = db.ExecContext(ctx,
+		`INSERT INTO results (task_id) VALUES ($1) ON CONFLICT (task_id) DO NOTHING`, taskID)
+	if err != nil {
+		return false, fmt.Errorf("ensure row: %w", err)
+	}
+
+	err = db.QueryRowContext(ctx, `
+		UPDATE results
+		SET tagging = $2::jsonb,
+		    filename = COALESCE(NULLIF($3, ''), filename),
+		    transcription = COALESCE(NULLIF($4, ''), transcription),
+		    updated_at = now(),
+		    status = CASE WHEN analysis IS NOT NULL THEN 'done' ELSE status END
+		WHERE task_id = $1
+		RETURNING (analysis IS NOT NULL AND tagging IS NOT NULL)
+	`, taskID, string(tagging), filename, transcription).Scan(&complete)
+	if err != nil {
+		return false, fmt.Errorf("update tagging: %w", err)
+	}
+	return complete, nil
+}
+
+// ===================== MAIN =====================
+
+func loadDotenv() {
+	path := os.Getenv("DOTENV_PATH")
+	if path == "" {
+		return
+	}
+	if err := godotenv.Overload(path); err != nil {
+		slog.Warn("dotenv load failed", "path", path, "error", err)
+		return
+	}
+	slog.Info("dotenv loaded", "path", path)
+}
+
+func main() {
+	loadDotenv()
+
+	amqpURL := getenv("RABBITMQ_URL", "amqp://guest:guest@localhost:5672/")
+	dbURL := os.Getenv("DATABASE_URL")
+	token := os.Getenv("YANDEX_API_KEY")
+	model := os.Getenv("YANDEX_MODEL")
+	inputQueue := getenv("TAGGING_QUEUE", "tagging")
+	finalQueue := getenv("FINAL_QUEUE", "final")
+
+	if token == "" {
+		slog.Error("YANDEX_API_KEY is required")
+		os.Exit(1)
+	}
+	if model == "" {
+		slog.Error("YANDEX_MODEL is required")
+		os.Exit(1)
+	}
+	slog.Info("config loaded", "worker", "tagging",
+		"yandex_token", tokenFingerprint(token), "model", model, "api_url", apiURL())
+
+	db := mustDB(dbURL)
+	defer db.Close()
+
+	checkCtx, checkCancel := context.WithTimeout(context.Background(), 90*time.Second)
+	if err := checkYandexAPI(checkCtx, model); err != nil {
+		checkCancel()
+		slog.Error("yandex api check failed — worker will not start", "worker", "tagging", "error", err)
+		os.Exit(1)
+	}
+	checkCancel()
+
+	ch := mustRabbit(amqpURL)
+
+	if _, err := ch.QueueDeclare(inputQueue, true, false, false, false, nil); err != nil {
+		slog.Error("declare queue failed", "queue", inputQueue, "error", err)
+		os.Exit(1)
+	}
+	if _, err := ch.QueueDeclare(finalQueue, true, false, false, false, nil); err != nil {
+		slog.Error("declare queue failed", "queue", finalQueue, "error", err)
+		os.Exit(1)
+	}
+	ch.Qos(1, 0, false)
+
+	msgs, err := ch.Consume(inputQueue, "", false, false, false, false, nil)
+	if err != nil {
+		slog.Error("consume failed", "error", err)
+		os.Exit(1)
+	}
+	slog.Info("worker started", "worker", "tagging", "queue", inputQueue, "model", model)
+
+	for d := range msgs {
+		taskStart := time.Now()
+		var task WorkerMessage
+		if err := json.Unmarshal(d.Body, &task); err != nil {
+			slog.Warn("bad message", "worker", "tagging", "delivery_tag", d.DeliveryTag,
+				"body_bytes", len(d.Body), "error", err)
+			d.Nack(false, false)
+			continue
+		}
+
+		transcriptionChars := utf8.RuneCountInString(task.Transcription)
+		slog.Info("message received", "worker", "tagging",
+			"task_id", task.TaskID,
+			"filename", task.Filename,
+			"delivery_tag", d.DeliveryTag,
+			"redelivered", d.Redelivered,
+			"body_bytes", len(d.Body),
+			"transcription_chars", transcriptionChars,
+			"llm_calls_expected", 1,
+		)
+		if d.Redelivered {
+			slog.Warn("redelivered message skipped — no llm call",
+				"worker", "tagging", "task_id", task.TaskID, "delivery_tag", d.DeliveryTag)
+			d.Nack(false, false)
+			continue
+		}
+
+		ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
+
+		result, err := classify(ctx, task.TaskID, model, task.Transcription)
+		if err != nil {
+			cancel()
+			slog.Warn("task failed, discarded",
+				"worker", "tagging", "task_id", task.TaskID,
+				"llm_calls", 1, "error", err)
+			d.Nack(false, false)
+			continue
+		}
+
+		tagJSON, _ := json.Marshal(result)
+		complete, err := saveTagging(ctx, db, task.TaskID, task.Filename, task.Transcription, tagJSON)
+		if err != nil {
+			cancel()
+			slog.Warn("db save failed, discarded",
+				"worker", "tagging", "task_id", task.TaskID, "error", err)
+			d.Nack(false, false)
+			continue
+		}
+
+		if complete {
+			notifyFinal(ctx, ch, db, finalQueue, task.TaskID, "tagging")
+			slog.Info("task complete", "worker", "tagging", "task_id", task.TaskID,
+				"was_last", "tagging", "L1", result.L1,
+				"llm_calls", 1, "duration_ms", time.Since(taskStart).Milliseconds())
+		} else {
+			slog.Info("task partial", "worker", "tagging", "task_id", task.TaskID,
+				"waiting_for", "analyse", "L1", result.L1,
+				"llm_calls", 1, "duration_ms", time.Since(taskStart).Milliseconds())
+		}
+		cancel()
+
+		d.Ack(false)
+	}
+}
+
+func getenv(k, d string) string {
+	if v := os.Getenv(k); v != "" {
+		return v
+	}
+	return d
+}
+
+func tokenFingerprint(token string) string {
+	if len(token) <= 12 {
+		return "***"
+	}
+	return token[:8] + "..." + token[len(token)-4:]
+}
+
+func truncate(s string, max int) string {
+	if len(s) <= max {
+		return s
+	}
+	return s[:max] + "..."
+}
+
+func loadFinalPayload(ctx context.Context, db *sql.DB, taskID string) ([]byte, error) {
+	var (
+		filename, transcription, status sql.NullString
+		analysis, tagging, metadata     []byte
+		createdAt, updatedAt            time.Time
+	)
+	err := db.QueryRowContext(ctx, `
+		SELECT filename, transcription, analysis, tagging, metadata, status, created_at, updated_at
+		FROM results WHERE task_id = $1
+	`, taskID).Scan(&filename, &transcription, &analysis, &tagging, &metadata, &status, &createdAt, &updatedAt)
+	if err != nil {
+		return nil, fmt.Errorf("load result: %w", err)
+	}
+
+	msg := map[string]any{
+		"task_id":    taskID,
+		"status":     status.String,
+		"created_at": createdAt,
+		"updated_at": updatedAt,
+	}
+	if filename.Valid {
+		msg["filename"] = filename.String
+	}
+	if transcription.Valid {
+		msg["transcription"] = transcription.String
+	}
+	if len(analysis) > 0 {
+		var v any
+		if err := json.Unmarshal(analysis, &v); err == nil {
+			msg["analysis"] = v
+		}
+	}
+	if len(tagging) > 0 {
+		var v any
+		if err := json.Unmarshal(tagging, &v); err == nil {
+			msg["tagging"] = v
+		}
+	}
+	if len(metadata) > 0 {
+		var meta map[string]any
+		if err := json.Unmarshal(metadata, &meta); err == nil {
+			for k, v := range meta {
+				msg[k] = v
+			}
+		}
+	}
+	return json.Marshal(msg)
+}
+
+func notifyFinal(ctx context.Context, ch *amqp.Channel, db *sql.DB, queue, taskID, worker string) {
+	body, err := loadFinalPayload(ctx, db, taskID)
+	if err != nil {
+		slog.Warn("load final payload failed", "worker", worker, "task_id", taskID, "error", err)
+		return
+	}
+	if err := ch.PublishWithContext(ctx, "", queue, false, false,
+		amqp.Publishing{
+			ContentType:  "application/json",
+			Body:         body,
+			DeliveryMode: amqp.Persistent,
+		}); err != nil {
+		slog.Warn("publish final failed", "worker", worker, "task_id", taskID, "error", err)
+		return
+	}
+	slog.Info("published final", "worker", worker, "task_id", taskID, "queue", queue, "body_bytes", len(body))
+	deleteProcessingFile(extractFilePath(body), taskID, worker)
+}
+
+func extractFilePath(body []byte) string {
+	var msg map[string]any
+	if err := json.Unmarshal(body, &msg); err != nil {
+		return ""
+	}
+	fp, _ := msg["file_path"].(string)
+	return fp
+}
+
+func deleteProcessingFile(filePath, taskID, worker string) {
+	if filePath == "" {
+		slog.Warn("processing file not deleted: no file_path", "worker", worker, "task_id", taskID)
+		return
+	}
+	if !strings.Contains(filePath, "/processing/") {
+		slog.Warn("processing file not deleted: path outside processing", "worker", worker, "task_id", taskID, "path", filePath)
+		return
+	}
+	if err := os.Remove(filePath); err != nil {
+		if os.IsNotExist(err) {
+			slog.Info("processing file already removed", "worker", worker, "task_id", taskID, "path", filePath)
+			return
+		}
+		slog.Warn("processing file delete failed", "worker", worker, "task_id", taskID, "path", filePath, "error", err)
+		return
+	}
+	slog.Info("processing file deleted", "worker", worker, "task_id", taskID, "path", filePath)
+}
+
+func mustDB(url string) *sql.DB {
+	db, err := sql.Open("pgx", url)
+	if err != nil {
+		slog.Error("db open failed", "error", err)
+		os.Exit(1)
+	}
+	db.SetMaxOpenConns(5)
+	time.Sleep(2 * time.Second) // дать Docker DNS зарегистрировать postgres
+	for i := 0; i < 60; i++ {
+		if err = db.Ping(); err == nil {
+			return db
+		}
+		if i < 5 || (i+1)%10 == 0 {
+			slog.Info("waiting for db", "attempt", i+1, "error", err)
+		}
+		time.Sleep(3 * time.Second)
+	}
+	slog.Error("db unreachable", "error", err)
+	os.Exit(1)
+	return nil
+}
+
+func mustRabbit(url string) *amqp.Channel {
+	var conn *amqp.Connection
+	var err error
+	for i := 0; i < 30; i++ {
+		conn, err = amqp.Dial(url)
+		if err == nil {
+			break
+		}
+		slog.Info("waiting for rabbit", "attempt", i+1, "error", err)
+		time.Sleep(2 * time.Second)
+	}
+	if err != nil {
+		slog.Error("rabbit unreachable", "error", err)
+		os.Exit(1)
+	}
+	ch, err := conn.Channel()
+	if err != nil {
+		slog.Error("rabbit channel failed", "error", err)
+		os.Exit(1)
+	}
+	return ch
+}
--- a/workers/tagging/go.mod
+++ b/workers/tagging/go.mod
@@ -0,0 +1,18 @@
+module github.com/yourorg/tagging
+
+go 1.22
+
+require (
+	github.com/jackc/pgx/v5 v5.5.5
+	github.com/joho/godotenv v1.5.1
+	github.com/rabbitmq/amqp091-go v1.9.0
+)
+
+require (
+	github.com/jackc/pgpassfile v1.0.0 // indirect
+	github.com/jackc/pgservicefile v0.0.0-20221227161230-091c0ba34f0a // indirect
+	github.com/jackc/puddle/v2 v2.2.1 // indirect
+	golang.org/x/crypto v0.17.0 // indirect
+	golang.org/x/sync v0.1.0 // indirect
+	golang.org/x/text v0.14.0 // indirect
+)
--- a/workers/tagging/go.sum
+++ b/workers/tagging/go.sum
@@ -0,0 +1,41 @@
+github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
+github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/jackc/pgpassfile v1.0.0 h1:/6Hmqy13Ss2zCq62VdNG8tM1wchn8zjSGOBJ6icpsIM=
+github.com/jackc/pgpassfile v1.0.0/go.mod h1:CEx0iS5ambNFdcRtxPj5JhEz+xB6uRky5eyVu/W2HEg=
+github.com/jackc/pgservicefile v0.0.0-20221227161230-091c0ba34f0a h1:bbPeKD0xmW/Y25WS6cokEszi5g+S0QxI/d45PkRi7Nk=
+github.com/jackc/pgservicefile v0.0.0-20221227161230-091c0ba34f0a/go.mod h1:5TJZWKEWniPve33vlWYSoGYefn3gLQRzjfDlhSJ9ZKM=
+github.com/jackc/pgx/v5 v5.5.5 h1:amBjrZVmksIdNjxGW/IiIMzxMKZFelXbUoPNb+8sjQw=
+github.com/jackc/pgx/v5 v5.5.5/go.mod h1:ez9gk+OAat140fv9ErkZDYFWmXLfV+++K0uAOiwgm1A=
+github.com/jackc/puddle/v2 v2.2.1 h1:RhxXJtFG022u4ibrCSMSiu5aOq1i77R3OHKNJj77OAk=
+github.com/jackc/puddle/v2 v2.2.1/go.mod h1:vriiEXHvEE654aYKXXjOvZM39qJ0q+azkZFrfEOc3H4=
+github.com/joho/godotenv v1.5.1 h1:7eLL/+HRGLY0ldzfGMeQkb7vMd0as4CfYvUVzLqw0N0=
+github.com/joho/godotenv v1.5.1/go.mod h1:f4LDr5Voq0i2e/R5DDNOoa2zzDfwtkZa6DnEwAbqwq4=
+github.com/kr/pretty v0.1.0/go.mod h1:dAy3ld7l9f0ibDNOQOHHMYYIIbhfbHSm3C4ZsoJORNo=
+github.com/kr/pty v1.1.1/go.mod h1:pFQYn66WHrOpPYNljwOMqo10TkYh1fy3cYio2l3bCsQ=
+github.com/kr/text v0.1.0/go.mod h1:4Jbv+DJW3UT/LiOwJeYQe1efqtUx/iVham/4vfdArNI=
+github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
+github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
+github.com/rabbitmq/amqp091-go v1.9.0 h1:qrQtyzB4H8BQgEuJwhmVQqVHB9O4+MNDJCCAcpc3Aoo=
+github.com/rabbitmq/amqp091-go v1.9.0/go.mod h1:+jPrT9iY2eLjRaMSRHUhc3z14E/l85kv/f+6luSD3pc=
+github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
+github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=
+github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=
+github.com/stretchr/testify v1.7.0/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
+github.com/stretchr/testify v1.7.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
+github.com/stretchr/testify v1.8.0/go.mod h1:yNjHg4UonilssWZ8iaSj1OCr/vHnekPRkoO+kdMU+MU=
+github.com/stretchr/testify v1.8.1 h1:w7B6lhMri9wdJUVmEZPGGhZzrYTPvgJArz7wNPgYKsk=
+github.com/stretchr/testify v1.8.1/go.mod h1:w2LPCIKwWwSfY2zedu0+kehJoqGctiVI29o6fzry7u4=
+go.uber.org/goleak v1.2.1 h1:NBol2c7O1ZokfZ0LEU9K6Whx/KnwvepVetCUhtKja4A=
+go.uber.org/goleak v1.2.1/go.mod h1:qlT2yGI9QafXHhZZLxlSuNsMw3FFLxBr+tBRlmO1xH4=
+golang.org/x/crypto v0.17.0 h1:r8bRNjWL3GshPW3gkd+RpvzWrZAwPS49OmTGZ/uhM4k=
+golang.org/x/crypto v0.17.0/go.mod h1:gCAAfMLgwOJRpTjQ2zCCt2OcSfYMTeZVSRtQlPC7Nq4=
+golang.org/x/sync v0.1.0 h1:wsuoTGHzEhffawBOhz5CYhcrV4IdKZbEyZjBMuTp12o=
+golang.org/x/sync v0.1.0/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
+golang.org/x/text v0.14.0 h1:ScX5w1eTa3QqT8oi6+ziP7dTV1S2+ALU0bI+0zXKWiQ=
+golang.org/x/text v0.14.0/go.mod h1:18ZOQIKpY8NJVqYksKHtTdi31H5itFRjB5/qKTNYzSU=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/check.v1 v1.0.0-20180628173108-788fd7840127/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
+gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
+gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
--- a/workers/transcribe/Dockerfile
+++ b/workers/transcribe/Dockerfile
@@ -0,0 +1,11 @@
+FROM golang:1.22-alpine AS build
+WORKDIR /src
+COPY go.mod go.sum* ./
+RUN go mod download
+COPY . .
+RUN CGO_ENABLED=0 go build -o /transcribe ./cmd/transcribe
+
+FROM alpine:3.19
+RUN apk add --no-cache ca-certificates
+COPY --from=build /transcribe /transcribe
+ENTRYPOINT ["/transcribe"]
--- a/workers/transcribe/cmd/transcribe/main.go
+++ b/workers/transcribe/cmd/transcribe/main.go
@@ -0,0 +1,64 @@
+package main
+
+import (
+	"context"
+	"log/slog"
+	"os"
+	"os/signal"
+	"syscall"
+	"time"
+
+	amqp "github.com/rabbitmq/amqp091-go"
+
+	"github.com/yourorg/transcribe/internal/config"
+	"github.com/yourorg/transcribe/internal/consumer"
+)
+
+func main() {
+	slog.SetDefault(slog.New(slog.NewJSONHandler(os.Stdout, &slog.HandlerOptions{Level: slog.LevelInfo})))
+
+	cfg := config.Load()
+	if cfg.NexaraAPIKey == "" {
+		slog.Error("NEXARA_API_KEY is required")
+		os.Exit(1)
+	}
+
+	ch := mustRabbit(cfg.RabbitURL)
+	cons, err := consumer.New(cfg, ch)
+	if err != nil {
+		slog.Error("consumer init failed", "error", err)
+		os.Exit(1)
+	}
+
+	ctx, stop := signal.NotifyContext(context.Background(), syscall.SIGINT, syscall.SIGTERM)
+	defer stop()
+
+	if err := cons.Run(ctx); err != nil && ctx.Err() == nil {
+		slog.Error("consumer stopped", "error", err)
+		os.Exit(1)
+	}
+	slog.Info("transcribe worker stopping")
+}
+
+func mustRabbit(url string) *amqp.Channel {
+	var conn *amqp.Connection
+	var err error
+	for i := 0; i < 30; i++ {
+		conn, err = amqp.Dial(url)
+		if err == nil {
+			break
+		}
+		slog.Info("waiting for rabbit", "attempt", i+1, "error", err)
+		time.Sleep(2 * time.Second)
+	}
+	if err != nil {
+		slog.Error("rabbit unreachable", "error", err)
+		os.Exit(1)
+	}
+	ch, err := conn.Channel()
+	if err != nil {
+		slog.Error("rabbit channel failed", "error", err)
+		os.Exit(1)
+	}
+	return ch
+}
--- a/workers/transcribe/configs/prompts.json
+++ b/workers/transcribe/configs/prompts.json
@@ -0,0 +1,23 @@
+[
+  {
+    "id": 1,
+    "id_section": 1,
+    "name": "behavioral",
+    "prompt": "Ты — строгий классификатор звонков.\n\nЗадача:\nПроанализируй диалог и оцени поведенческие критерии.\n\nКритерии:\n1. Приветствие\n2. Инициативность (выявление цели, попытка развить разговор)\n3. Уточнил, остались ли вопросы\n4. Прощание\n\nИнструкция:\nДля каждого критерия:\n- определи наличие\n- найди ДОСЛОВНУЮ цитату\n- оцени confidence (0.0–1.0)\n\nФормат ответа (строго JSON):\n\n{\n  \"greeting\": {\n    \"value\": true/false,\n    \"evidence\": \"цитата или null\",\n    \"confidence\": number\n  },\n  \"initiative\": {\n    \"value\": true/false,\n    \"evidence\": \"цитата или null\",\n    \"confidence\": number\n  },\n  \"questions_check\": {\n    \"value\": true/false,\n    \"evidence\": \"цитата или null\",\n    \"confidence\": number\n  },\n  \"closing\": {\n    \"value\": true/false,\n    \"evidence\": \"цитата или null\",\n    \"confidence\": number\n  }\n}\n\nЖЁСТКИЕ ПРАВИЛА:\n- каждый критерий оценивается независимо\n- не додумывать\n- если нет → value=false, evidence=null, confidence=0.0\n- evidence должен подтверждать вывод",
+    "dt_create": "2026-06-09T09:00:00.000000"
+  },
+  {
+    "id": 2,
+    "id_section": 1,
+    "name": "client_data",
+    "prompt": "Ты — строгий классификатор звонков.\n\nЗадача:\nОпредели, какие данные о клиенте были получены.\n\nКритерии:\n1. Первый ли раз обращается\n2. Указан ли город клиента\n3. Тип клиента (физ/юр)\n4. Получены ли контакты\n5. Источник (откуда узнали)\n\nФормат ответа (строго JSON):\n\n{\n  \"first_time\": {\n    \"value\": true/false,\n    \"evidence\": \"цитата или null\",\n    \"confidence\": number\n  },\n  \"client_city\": {\n    \"value\": true/false,\n    \"city\": \"строка или null\",\n    \"evidence\": \"цитата или null\",\n    \"confidence\": number\n  },\n  \"client_type\": {\n    \"value\": true/false,\n    \"type\": \"physical|legal|null\",\n    \"evidence\": \"цитата или null\",\n    \"confidence\": number\n  },\n  \"contacts\": {\n    \"value\": true/false,\n    \"evidence\": \"цитата или null\",\n    \"confidence\": number\n  },\n  \"source\": {\n    \"value\": true/false,\n    \"evidence\": \"цитата или null\",\n    \"confidence\": number\n  }\n}\n\nЖЁСТКИЕ ПРАВИЛА:\n- city/type только если явно сказано\n- не додумывать\n- если нет → value=false, evidence=null, confidence=0.0",
+    "dt_create": "2026-06-09T09:00:00.000000"
+  },
+  {
+    "id": 3,
+    "id_section": 1,
+    "name": "cargo_data",
+    "prompt": "Ты — строгий классификатор логистических данных.\n\nКритерии:\n1. Характер груза\n2. Параметры груза (вес, объем, размеры)\n3. Стоимость груза\n\nФормат ответа (строго JSON):\n\n{\n  \"cargo_type\": {\n    \"value\": true/false,\n    \"type\": \"строка или null\",\n    \"evidence\": \"цитата или null\",\n    \"confidence\": number\n  },\n  \"cargo_params\": {\n    \"value\": true/false,\n    \"params\": \"строка или null\",\n    \"evidence\": \"цитата или null\",\n    \"confidence\": number\n  },\n  \"cargo_value\": {\n    \"value\": true/false,\n    \"amount\": \"строка или null\",\n    \"evidence\": \"цитата или null\",\n    \"confidence\": number\n  }\n}\n\nЖЁСТКИЕ ПРАВИЛА:\n- только явные данные\n- числа/параметры должны быть в evidence\n- если нет → value=false, evidence=null, confidence=0.0",
+    "dt_create": "2026-06-09T09:00:00.000000"
+  }
+]
--- a/workers/transcribe/go.mod
+++ b/workers/transcribe/go.mod
@@ -0,0 +1,5 @@
+module github.com/yourorg/transcribe
+
+go 1.22
+
+require github.com/rabbitmq/amqp091-go v1.9.0
--- a/workers/transcribe/go.sum
+++ b/workers/transcribe/go.sum
@@ -0,0 +1,18 @@
+github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/kr/pretty v0.1.0/go.mod h1:dAy3ld7l9f0ibDNOQOHHMYYIIbhfbHSm3C4ZsoJORNo=
+github.com/kr/pty v1.1.1/go.mod h1:pFQYn66WHrOpPYNljwOMqo10TkYh1fy3cYio2l3bCsQ=
+github.com/kr/text v0.1.0/go.mod h1:4Jbv+DJW3UT/LiOwJeYQe1efqtUx/iVham/4vfdArNI=
+github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
+github.com/rabbitmq/amqp091-go v1.9.0 h1:qrQtyzB4H8BQgEuJwhmVQqVHB9O4+MNDJCCAcpc3Aoo=
+github.com/rabbitmq/amqp091-go v1.9.0/go.mod h1:+jPrT9iY2eLjRaMSRHUhc3z14E/l85kv/f+6luSD3pc=
+github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
+github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=
+github.com/stretchr/testify v1.7.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
+github.com/stretchr/testify v1.8.0/go.mod h1:yNjHg4UonilssWZ8iaSj1OCr/vHnekPRkoO+kdMU+MU=
+go.uber.org/goleak v1.2.1 h1:NBol2c7O1ZokfZ0LEU9K6Whx/KnwvepVetCUhtKja4A=
+go.uber.org/goleak v1.2.1/go.mod h1:qlT2yGI9QafXHhZZLxlSuNsMw3FFLxBr+tBRlmO1xH4=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/check.v1 v1.0.0-20180628173108-788fd7840127/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
+gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
--- a/workers/transcribe/internal/config/config.go
+++ b/workers/transcribe/internal/config/config.go
@@ -0,0 +1,78 @@
+package config
+
+import (
+	"os"
+	"strconv"
+	"time"
+)
+
+type Config struct {
+	RabbitURL       string
+	InputQueue      string
+	OutputExchange  string
+	AnalyseQueue    string
+	TaggingQueue    string
+	InputExchange   string
+	InputRoutingKey string
+	Prefetch        int
+
+	NexaraBaseURL string
+	NexaraAPIKey  string
+	NexaraModel   string
+	NexaraTimeout time.Duration
+
+	PromptsSource  string
+	PromptsFile    string
+	PromptsBaseURL string
+	PromptsAPIKey  string
+	PromptsSection int
+}
+
+func Load() Config {
+	return Config{
+		RabbitURL:       getEnv("RABBITMQ_URL", "amqp://guest:guest@localhost:5672/"),
+		InputQueue:      getEnv("INPUT_QUEUE", "transcribe.tasks"),
+		OutputExchange:  getEnv("OUTPUT_EXCHANGE", "transcription_done"),
+		AnalyseQueue:    getEnv("ANALYSE_QUEUE", "analyse"),
+		TaggingQueue:    getEnv("TAGGING_QUEUE", "tagging"),
+		InputExchange:   getEnv("RABBITMQ_EXCHANGE", "audio_pipeline"),
+		InputRoutingKey: getEnv("RABBITMQ_ROUTING_KEY", "audio.new"),
+		Prefetch:        getInt("PREFETCH", 1),
+
+		NexaraBaseURL: getEnv("NEXARA_BASE_URL", "https://api.nexara.ru"),
+		NexaraAPIKey:  os.Getenv("NEXARA_API_KEY"),
+		NexaraModel:   getEnv("NEXARA_MODEL", "whisper-1"),
+		NexaraTimeout: getDuration("NEXARA_TIMEOUT", 10*time.Minute),
+
+		PromptsSource:  getEnv("PROMPTS_SOURCE", "static"),
+		PromptsFile:    getEnv("PROMPTS_FILE", "/app/configs/prompts.json"),
+		PromptsBaseURL: os.Getenv("PROMPTS_BASE_URL"),
+		PromptsAPIKey:  os.Getenv("PROMPTS_API_KEY"),
+		PromptsSection: getInt("PROMPTS_SECTION", 1),
+	}
+}
+
+func getEnv(key, def string) string {
+	if v := os.Getenv(key); v != "" {
+		return v
+	}
+	return def
+}
+
+func getInt(key string, def int) int {
+	if v := os.Getenv(key); v != "" {
+		if i, err := strconv.Atoi(v); err == nil {
+			return i
+		}
+	}
+	return def
+}
+
+func getDuration(key string, def time.Duration) time.Duration {
+	if v := os.Getenv(key); v != "" {
+		if d, err := time.ParseDuration(v); err == nil {
+			return d
+		}
+	}
+	return def
+}
--- a/workers/transcribe/internal/consumer/consumer.go
+++ b/workers/transcribe/internal/consumer/consumer.go
@@ -0,0 +1,172 @@
+package consumer
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"log/slog"
+	"time"
+
+	amqp "github.com/rabbitmq/amqp091-go"
+
+	"github.com/yourorg/transcribe/internal/config"
+	"github.com/yourorg/transcribe/internal/models"
+	"github.com/yourorg/transcribe/internal/nexara"
+	"github.com/yourorg/transcribe/internal/prompts"
+)
+
+type Consumer struct {
+	cfg     config.Config
+	ch      *amqp.Channel
+	nexara  *nexara.Client
+	prompts *prompts.Loader
+}
+
+func New(cfg config.Config, ch *amqp.Channel) (*Consumer, error) {
+	if err := setupTopology(ch, cfg); err != nil {
+		return nil, err
+	}
+	return &Consumer{
+		cfg:     cfg,
+		ch:      ch,
+		nexara:  nexara.New(cfg.NexaraBaseURL, cfg.NexaraAPIKey, cfg.NexaraModel, cfg.NexaraTimeout),
+		prompts: prompts.New(cfg.PromptsSource, cfg.PromptsFile, cfg.PromptsBaseURL, cfg.PromptsAPIKey, cfg.PromptsSection),
+	}, nil
+}
+
+func setupTopology(ch *amqp.Channel, cfg config.Config) error {
+	if err := ch.ExchangeDeclare("dlx", "direct", true, false, false, false, nil); err != nil {
+		return fmt.Errorf("declare dlx: %w", err)
+	}
+	if err := ch.ExchangeDeclare(cfg.InputExchange, "direct", true, false, false, false, nil); err != nil {
+		return fmt.Errorf("declare input exchange: %w", err)
+	}
+	if err := ch.ExchangeDeclare(cfg.OutputExchange, "fanout", true, false, false, false, nil); err != nil {
+		return fmt.Errorf("declare output exchange: %w", err)
+	}
+
+	dlqArgs := amqp.Table{
+		"x-dead-letter-exchange":    "dlx",
+		"x-dead-letter-routing-key": cfg.InputQueue + ".failed",
+	}
+	if _, err := ch.QueueDeclare(cfg.InputQueue, true, false, false, false, dlqArgs); err != nil {
+		return fmt.Errorf("declare input queue: %w", err)
+	}
+	if _, err := ch.QueueDeclare(cfg.InputQueue+".failed", true, false, false, false, nil); err != nil {
+		return fmt.Errorf("declare dlq: %w", err)
+	}
+	if err := ch.QueueBind(cfg.InputQueue+".failed", cfg.InputQueue+".failed", "dlx", false, nil); err != nil {
+		return fmt.Errorf("bind dlq: %w", err)
+	}
+	if err := ch.QueueBind(cfg.InputQueue, cfg.InputRoutingKey, cfg.InputExchange, false, nil); err != nil {
+		return fmt.Errorf("bind input queue: %w", err)
+	}
+
+	for _, q := range []string{cfg.AnalyseQueue, cfg.TaggingQueue} {
+		if _, err := ch.QueueDeclare(q, true, false, false, false, nil); err != nil {
+			return fmt.Errorf("declare queue %s: %w", q, err)
+		}
+		if err := ch.QueueBind(q, "", cfg.OutputExchange, false, nil); err != nil {
+			return fmt.Errorf("bind queue %s: %w", q, err)
+		}
+	}
+
+	return ch.Qos(cfg.Prefetch, 0, false)
+}
+
+func (c *Consumer) Run(ctx context.Context) error {
+	if err := c.ch.Confirm(false); err != nil {
+		return fmt.Errorf("confirm mode: %w", err)
+	}
+
+	msgs, err := c.ch.Consume(c.cfg.InputQueue, "", false, false, false, false, nil)
+	if err != nil {
+		return err
+	}
+
+	slog.Info("transcribe worker started", "queue", c.cfg.InputQueue, "output_exchange", c.cfg.OutputExchange)
+
+	for {
+		select {
+		case <-ctx.Done():
+			return nil
+		case d, ok := <-msgs:
+			if !ok {
+				return fmt.Errorf("delivery channel closed")
+			}
+			c.handle(ctx, d)
+		}
+	}
+}
+
+func (c *Consumer) handle(ctx context.Context, d amqp.Delivery) {
+	var task models.AudioTask
+	if err := json.Unmarshal(d.Body, &task); err != nil {
+		slog.Warn("bad message", "delivery_tag", d.DeliveryTag, "error", err)
+		_ = d.Nack(false, false)
+		return
+	}
+
+	slog.Info("message received", "task_id", task.TaskID, "file_path", task.FilePath, "filename", task.Filename)
+
+	txCtx, cancel := context.WithTimeout(ctx, c.cfg.NexaraTimeout+30*time.Second)
+	defer cancel()
+
+	text, lang, segments, err := c.nexara.TranscribeFile(txCtx, task.FilePath)
+	if err != nil {
+		slog.Warn("transcription failed", "task_id", task.TaskID, "error", err)
+		_ = d.Nack(false, false)
+		return
+	}
+
+	promptList, err := c.prompts.Load(txCtx)
+	if err != nil {
+		slog.Warn("prompts load failed", "task_id", task.TaskID, "error", err)
+		_ = d.Nack(false, false)
+		return
+	}
+
+	result := models.TranscriptionResult{
+		TaskID:        task.TaskID,
+		Filename:      task.Filename,
+		FilePath:      task.FilePath,
+		Transcription: text,
+		Language:      lang,
+		Segments:      segments,
+		Prompts:       promptList,
+		TranscribedAt: time.Now().Unix(),
+	}
+
+	body, err := json.Marshal(result)
+	if err != nil {
+		slog.Warn("marshal failed", "task_id", task.TaskID, "error", err)
+		_ = d.Nack(false, false)
+		return
+	}
+
+	confirms := c.ch.NotifyPublish(make(chan amqp.Confirmation, 1))
+	if err := c.ch.PublishWithContext(txCtx, c.cfg.OutputExchange, "", false, false, amqp.Publishing{
+		ContentType:  "application/json",
+		Body:         body,
+		DeliveryMode: amqp.Persistent,
+	}); err != nil {
+		slog.Warn("publish failed, requeue", "task_id", task.TaskID, "error", err)
+		_ = d.Nack(false, true)
+		return
+	}
+	select {
+	case confirm := <-confirms:
+		if !confirm.Ack {
+			slog.Warn("publish not confirmed, requeue", "task_id", task.TaskID)
+			_ = d.Nack(false, true)
+			return
+		}
+	case <-txCtx.Done():
+		slog.Warn("publish timeout, requeue", "task_id", task.TaskID)
+		_ = d.Nack(false, true)
+		return
+	}
+
+	slog.Info("transcribed", "task_id", task.TaskID, "language", lang, "chars", len(text), "segments", len(segments), "prompts", len(promptList))
+	_ = d.Ack(false)
+}
--- a/workers/transcribe/internal/models/models.go
+++ b/workers/transcribe/internal/models/models.go
@@ -0,0 +1,34 @@
+package models
+
+type AudioTask struct {
+	TaskID    string `json:"task_id"`
+	FilePath  string `json:"file_path"`
+	Filename  string `json:"filename"`
+	Size      int64  `json:"size"`
+	CreatedAt int64  `json:"created_at"`
+}
+
+type Segment struct {
+	Start float64 `json:"start"`
+	End   float64 `json:"end"`
+	Text  string  `json:"text"`
+}
+
+type Prompt struct {
+	ID        int    `json:"id"`
+	IDSection int    `json:"id_section"`
+	Name      string `json:"name"`
+	Prompt    string `json:"prompt"`
+	DtCreate  string `json:"dt_create"`
+}
+
+type TranscriptionResult struct {
+	TaskID        string    `json:"task_id"`
+	Filename      string    `json:"filename"`
+	FilePath      string    `json:"file_path"`
+	Transcription string    `json:"transcription"`
+	Language      string    `json:"language"`
+	Segments      []Segment `json:"segments,omitempty"`
+	Prompts       []Prompt  `json:"prompts"`
+	TranscribedAt int64     `json:"transcribed_at"`
+}
--- a/workers/transcribe/internal/nexara/nexara.go
+++ b/workers/transcribe/internal/nexara/nexara.go
@@ -0,0 +1,117 @@
+package nexara
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"mime/multipart"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+
+	"github.com/yourorg/transcribe/internal/models"
+)
+
+type Client struct {
+	apiURL     string
+	apiKey     string
+	model      string
+	httpClient *http.Client
+}
+
+func New(baseURL, apiKey, model string, timeout time.Duration) *Client {
+	baseURL = strings.TrimRight(baseURL, "/")
+	return &Client{
+		apiURL: baseURL + "/api/v1/audio/transcriptions",
+		apiKey: apiKey,
+		model:  model,
+		httpClient: &http.Client{
+			Timeout: timeout,
+		},
+	}
+}
+
+func (c *Client) TranscribeFile(ctx context.Context, path string) (text, language string, segments []models.Segment, err error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return "", "", nil, fmt.Errorf("open file: %w", err)
+	}
+	defer f.Close()
+
+	body := &bytes.Buffer{}
+	writer := multipart.NewWriter(body)
+	part, err := writer.CreateFormFile("file", filepath.Base(path))
+	if err != nil {
+		return "", "", nil, err
+	}
+	if _, err := io.Copy(part, f); err != nil {
+		return "", "", nil, err
+	}
+	if c.model != "" {
+		if err := writer.WriteField("model", c.model); err != nil {
+			return "", "", nil, err
+		}
+	}
+	if err := writer.WriteField("response_format", "json"); err != nil {
+		return "", "", nil, err
+	}
+	if err := writer.Close(); err != nil {
+		return "", "", nil, err
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, c.apiURL, body)
+	if err != nil {
+		return "", "", nil, err
+	}
+	req.Header.Set("Content-Type", writer.FormDataContentType())
+	req.Header.Set("Authorization", "Bearer "+c.apiKey)
+
+	resp, err := c.httpClient.Do(req)
+	if err != nil {
+		return "", "", nil, fmt.Errorf("request: %w", err)
+	}
+	defer resp.Body.Close()
+
+	respBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return "", "", nil, err
+	}
+	if resp.StatusCode != http.StatusOK {
+		return "", "", nil, fmt.Errorf("status %d: %s", resp.StatusCode, string(respBody))
+	}
+
+	var raw map[string]any
+	if err := json.Unmarshal(respBody, &raw); err != nil {
+		return "", "", nil, fmt.Errorf("parse: %w", err)
+	}
+	if t, ok := raw["text"].(string); ok {
+		text = t
+	}
+	if lang, ok := raw["language"].(string); ok {
+		language = lang
+	}
+	if segs, ok := raw["segments"].([]any); ok {
+		for _, s := range segs {
+			m, ok := s.(map[string]any)
+			if !ok {
+				continue
+			}
+			var seg models.Segment
+			if v, ok := m["start"].(float64); ok {
+				seg.Start = v
+			}
+			if v, ok := m["end"].(float64); ok {
+				seg.End = v
+			}
+			if v, ok := m["text"].(string); ok {
+				seg.Text = v
+			}
+			segments = append(segments, seg)
+		}
+	}
+	return text, language, segments, nil
+}
--- a/workers/transcribe/internal/prompts/prompts.go
+++ b/workers/transcribe/internal/prompts/prompts.go
@@ -0,0 +1,100 @@
+package prompts
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/yourorg/transcribe/internal/models"
+)
+
+type Loader struct {
+	source    string
+	filePath  string
+	baseURL   string
+	apiKey    string
+	sectionID int
+	client    *http.Client
+}
+
+func New(source, filePath, baseURL, apiKey string, sectionID int) *Loader {
+	return &Loader{
+		source:    source,
+		filePath:  filePath,
+		baseURL:   strings.TrimRight(baseURL, "/"),
+		apiKey:    apiKey,
+		sectionID: sectionID,
+		client:    &http.Client{Timeout: 30 * time.Second},
+	}
+}
+
+func (l *Loader) Load(ctx context.Context) ([]models.Prompt, error) {
+	switch strings.ToLower(l.source) {
+	case "http":
+		return l.loadHTTP(ctx)
+	default:
+		return l.loadStatic()
+	}
+}
+
+func (l *Loader) loadStatic() ([]models.Prompt, error) {
+	data, err := os.ReadFile(l.filePath)
+	if err != nil {
+		return nil, fmt.Errorf("read prompts file: %w", err)
+	}
+	var prompts []models.Prompt
+	if err := json.Unmarshal(data, &prompts); err != nil {
+		return nil, fmt.Errorf("parse prompts file: %w", err)
+	}
+	return filterSection(prompts, l.sectionID), nil
+}
+
+func (l *Loader) loadHTTP(ctx context.Context) ([]models.Prompt, error) {
+	if l.baseURL == "" {
+		return nil, fmt.Errorf("PROMPTS_BASE_URL is required for http source")
+	}
+	url := fmt.Sprintf("%s/metrics/?id_section=%s", l.baseURL, strconv.Itoa(l.sectionID))
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
+	if err != nil {
+		return nil, err
+	}
+	if l.apiKey != "" {
+		req.Header.Set("Authorization", "Bearer "+l.apiKey)
+	}
+	resp, err := l.client.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close()
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, err
+	}
+	if resp.StatusCode != http.StatusOK {
+		return nil, fmt.Errorf("prompts api status %d: %s", resp.StatusCode, string(body))
+	}
+	var prompts []models.Prompt
+	if err := json.Unmarshal(body, &prompts); err != nil {
+		return nil, fmt.Errorf("parse prompts response: %w", err)
+	}
+	return filterSection(prompts, l.sectionID), nil
+}
+
+func filterSection(prompts []models.Prompt, sectionID int) []models.Prompt {
+	if sectionID <= 0 {
+		return prompts
+	}
+	out := make([]models.Prompt, 0, len(prompts))
+	for _, p := range prompts {
+		if p.IDSection == sectionID {
+			out = append(out, p)
+		}
+	}
+	return out
+}