Проект 3: RAG Chatbot

🎯 Цель

Полный production-ready RAG (Retrieval Augmented Generation) chatbot. Multilingual для узбекских документов, полезный в локальном контексте AI assistant.

Architecture

┌────────────────┐
│  Web / Mobile  │
│  Telegram bot  │
└────────┬───────┘
         │
         ▼
┌────────────────────┐
│   FastAPI + WS     │
│   (Streaming)      │
└──────┬─────────────┘
       │
       ▼
┌──────────────────────────────────┐
│   RAG Pipeline                   │
│  ┌──────────┐  ┌──────────────┐ │
│  │  Query   │─>│  Embedding   │ │
│  │  Routing │  │  (OpenAI)    │ │
│  └──────────┘  └──────┬───────┘ │
│                       ▼          │
│  ┌─────────────────────────────┐ │
│  │  Qdrant Vector Search       │ │
│  │  + BM25 Hybrid + Rerank     │ │
│  └──────────┬──────────────────┘ │
│             ▼                    │
│  ┌─────────────────────────────┐ │
│  │  LLM (Claude / GPT)         │ │
│  │  + Citation                 │ │
│  └──────────┬──────────────────┘ │
└─────────────┼────────────────────┘
              │
              ▼
       ┌──────────────┐
       │  Postgres    │
       │  - History   │
       │  - Feedback  │
       └──────────────┘
              │
              ▼
       ┌──────────────┐
       │  Langfuse    │
       │  Observation │
       └──────────────┘

Tech Stack

Required

**Backend:**FastAPI (streaming + WebSocket)
**LLM:**OpenAI или Anthropic
**Vector DB:**Qdrant (или ChromaDB)
**Embeddings:**OpenAI text-embedding-3-small
**Framework:**LlamaIndex или raw API
**Frontend:**Streamlit или Next.js
**Container:**Docker + docker-compose

Nice to have

**Reranking:**Cross-encoder (BAAI)
**Observability:**Langfuse
**Telegram bot:**aiogram
**Cache:**Redis
**Authentication:**JWT

Features

MVP (1-я неделя)

Document ingestion (PDF, URL, MD)
Chunking + embeddings
Qdrant collection
FastAPI /chat endpoint
LLM API integration
Citation (basic)
Streamlit UI
Docker

V2 (2-я неделя)

Multi-source ingestion (PDF + URL + Notion)
Hybrid search (vector + BM25)
Reranking
Streaming responses (SSE)
Postgres conversation history
Multi-turn context
Tests
CI/CD

V3 (3-я неделя)

Telegram bot integration
Multi-language (uz/ru/en)
Langfuse observability
Feedback collection
RAGAS evaluation
Cloud deployment
Blog post

API spec

`POST /ingest`

# PDF upload
curl -X POST -F "file=@doc.pdf" -F "metadata={\"source\":\"law\"}" http://api/ingest

# URL
curl -X POST -d '{"url":"https://lex.uz/...","metadata":{}}' http://api/ingest

{
    "task_id": "uuid",
    "chunks_added": 234
}

`POST /chat`

// Request
{
    "message": "Какова максимальная продолжительность рабочей недели в Узбекистане?",
    "session_id": "uuid",
    "user_id": "user_123",
    "language": "uz"
}

// Response
{
    "answer": "Согласно статье 122 Трудового кодекса Узбекистана, рабочее время не должно превышать 40 часов в неделю [Source 1].",
    "sources": [
        {
            "text": "Обычная продолжительность рабочего времени не превышает 40 часов в неделю...",
            "document": "Трудовой кодекс",
            "section": "Статья 122",
            "url": "https://lex.uz/...#122",
            "score": 0.89
        }
    ],
    "session_id": "uuid",
    "model": "claude-sonnet-4-6",
    "tokens_used": 1245,
    "cost_usd": 0.003,
    "latency_ms": 1230
}

`POST /chat/stream` (SSE)

data: {"type": "sources", "data": [...]}
data: {"type": "token", "text": "Узбекистан "}
data: {"type": "token", "text": "Трудовой "}
...
data: {"type": "done", "total_tokens": 1245}

`POST /feedback`

{
    "session_id": "uuid",
    "message_id": "uuid",
    "rating": "thumbs_up",  // or thumbs_down
    "comment": "Точный ответ"
}

`GET /sessions/{user_id}`

Conversation history

`POST /telegram-webhook`

Telegram bot integration

Project structure

rag-chatbot/
├── README.md
├── docker-compose.yml
├── Dockerfile
├── .github/workflows/
├── src/
│   ├── api/
│   │   ├── main.py
│   │   ├── routes/
│   │   │   ├── chat.py
│   │   │   ├── ingest.py
│   │   │   ├── feedback.py
│   │   │   └── telegram.py
│   │   └── schemas.py
│   ├── core/
│   │   ├── config.py
│   │   └── prompts.py
│   ├── rag/
│   │   ├── ingestion.py
│   │   ├── retrieval.py
│   │   ├── reranking.py
│   │   ├── generation.py
│   │   └── pipeline.py             # full RAG
│   ├── llm/
│   │   ├── openai_client.py
│   │   └── anthropic_client.py
│   ├── vectordb/
│   │   └── qdrant_client.py
│   ├── data/
│   │   └── models.py               # Postgres
│   └── integrations/
│       └── telegram_bot.py
├── tests/
├── evaluation/
│   ├── test_set.json               # 100 Q&A pairs
│   └── ragas_eval.py
├── frontend/
│   └── streamlit_app.py
├── data/
│   └── documents/                  # source files
├── prompts/
│   └── system_v1.md
└── pyproject.toml

План реализации (3 недели)

Неделя 1 — MVP RAG

Day 1-2: Source documents collection + preparation
Day 3: Ingestion pipeline (chunking, embeddings, Qdrant)
Day 4: Basic RAG pipeline (retrieve → LLM → respond)
Day 5: FastAPI endpoint
Day 6: Streamlit UI
Day 7: Docker + GitHub

Неделя 2 — Advanced RAG

Day 8: Multi-source ingestion (PDF, URL, MD)
Day 9: Hybrid search (Qdrant native)
Day 10: Reranking (cross-encoder)
Day 11: Streaming SSE
Day 12: Postgres history + multi-turn
Day 13: Tests
Day 14: CI/CD

Неделя 3 — Production + Evaluation

Day 15: Telegram bot
Day 16: Multi-language support
Day 17: Langfuse observability
Day 18: Feedback + RAGAS evaluation
Day 19: Cloud deployment
Day 20: Demo video
Day 21: Blog post + LinkedIn

Success metrics

RAG quality (RAGAS metrics)

Faithfulness:> 0.85
Answer Relevancy:> 0.85
Context Precision:> 0.80
Context Recall:> 0.80

Performance

Retrieval latency:< 500ms
End-to-end latency:< 3s (non-streaming), TTFT < 1s (streaming)
Cost per query:< $0.01

User satisfaction

Thumbs up rate:> 75%
**Session retention:**users come back

Ресурсы

LlamaIndex docs — docs.llamaindex.ai
Qdrant tutorials — qdrant.tech/documentation
OpenAI Cookbook RAG — examples
Langfuse docs — observability
RAGAS docs — evaluation
Anthropic prompt caching — cost optimization

Bonus features

Multi-modal RAG — text + images (PDFs with charts)
Agentic RAG — LLM tools (search, calculator, DB)
Auto-evaluation — model judging model
Custom embedding model — domain-specific
Hybrid retrieval — vector + BM25 + knowledge graph
Multi-language search — query in EN, docs in UZ
Voice interface — Whisper STT + TTS
Mobile app — React Native

✅ Submission checklist

100+ документов ingestion
Hybrid search + reranking
Streaming responses
Multi-turn conversation
Citation with source links
Telegram bot working
Multi-language (минимум 2 языка)
RAGAS evaluation report
Langfuse dashboard
Streamlit UI live
Demo video
Blog post

Завершили? Проект 4: MLOps Pipeline — самый большой и самый важный проект.

Use case	Источник	Сложность
Chatbot по Конституции/УК Узбекистана	lex.uz	⭐⭐⭐⭐
Технический documentation bot	GitHub repo docs	⭐⭐⭐
Customer support bot	FAQ + product docs	⭐⭐⭐
HR / Internal docs	Notion / Confluence	⭐⭐⭐
Wikipedia chatbot (узбекский)	uz.wikipedia.org	⭐⭐⭐⭐
Medical knowledge base	Public medical docs	⭐⭐⭐⭐⭐
Legal advice bot	lex.uz + qonun.uz	⭐⭐⭐⭐⭐

Backend to ML: 6 Oylik Roadmap