Apppublic

ambarish0221/AskMyDoc

sourceHugging Facemitupdated 4mo agoView on Hugging Face

0likes

App README

<h1 align="center">AskMyDoc — AI Research Agent</h1>

<p align="center"> A production-grade <strong>single AI agent with RAG backbone</strong> — upload any document and converse with an autonomous agent that retrieves from your documents and the web in real time. </p>

Architecture

Browser (Next.js 15 + Tailwind)
        │  SSE stream (tokens + agent steps)
        ▼
FastAPI (port 8000)
  ├── Input guardrails       ← prompt injection detection, length limits
  ├── Session memory (20-turn history per UUID)
  ├── Rate limiting (slowapi, 20/min)
  ├── GET  /health            ← live status
  ├── POST /ingest            ← upload PDF or URL
  ├── GET  /documents         ← list ingested docs
  ├── DELETE /documents/{id}  ← remove single doc
  ├── POST /chat              ← SSE agent stream
  └── POST /evaluate          ← embedding-based metrics
        │
        ▼
 LangGraph ReAct Agent
   ├── rag_search  ─────► HyDE rewrite → Qdrant (BM25 + dense → cross-encoder re-rank top-6)
   │                          └── sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
   │                          └── cross-encoder/ms-marco-MiniLM-L-6-v2
   └── web_search  ─────► Tavily API
        │
        ▼
   LLM (OpenAI: gpt-4o-mini / gpt-4o / gpt-3.5-turbo)
      OR (HuggingFace: Mistral-7B / Zephyr-7B / Llama-3)

Project Structure

RAG-Agent-AskMyDoc/
├── docker-compose.yml
├── env.example
├── backend/
│   ├── Dockerfile
│   ├── requirements.txt
│   ├── main.py                   # FastAPI: all endpoints + session store
│   └── src/
│       ├── rag/
│       │   ├── loader.py         # PDF + URL document loaders
│       │   ├── chunkers.py       # 4 chunking strategies
│       ├── vector_store.py   # Qdrant client + multi-doc registry (multilingual embeddings)
│       │   ├── reranker.py       # Cross-encoder re-ranking
│       │   └── guardrails.py     # Input validation + output quality checks
│       └── agent/
│           ├── tools.py          # rag_search (HyDE + hybrid) + web_search tools
│           └── graph.py          # LangGraph ReAct agent (model selector)
└── frontend/
    ├── Dockerfile
    ├── app/
    │   ├── layout.tsx
    │   ├── page.tsx              # Layout: health banner + model selector + sidebar
    │   └── globals.css
    └── components/
        ├── ChatWindow.tsx        # SSE consumer + session ID management
        ├── MessageBubble.tsx     # Markdown + citations + eval panel
        ├── DocumentPanel.tsx     # Upload + doc list + per-doc delete
        ├── AgentTrace.tsx        # Collapsible tool call viewer
        ├── SourceCitation.tsx    # [Chunk N | file | strategy] badge parser
        └── EvalPanel.tsx         # answer_relevancy / faithfulness / context_recall bars

Quickstart (Docker Compose — recommended)

bash

git clone https://github.com/apatha32/RAG-Agent-AskMyDoc.git
cd RAG-Agent-AskMyDoc

cp env.example .env
# Edit .env — add OPENAI_API_KEY and/or HF_TOKEN

docker compose up --build

Open http://localhost:3000 — that's it.

Local Development (without Docker)

Backend

bash

cd backend
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp ../env.example .env   # add your keys
uvicorn main:app --reload --port 8000

No Qdrant running? The backend falls back to an in-memory Qdrant client automatically.

Frontend

bash

cd frontend
npm install
NEXT_PUBLIC_API_URL=http://localhost:8000 npm run dev

Open http://localhost:3000.

Environment Variables

Variable	Required	Description
`OPENAI_API_KEY`	For OpenAI models	platform.openai.com/api-keys
`HF_TOKEN`	For HuggingFace models	huggingface.co/settings/tokens
`TAVILY_API_KEY`	Optional — web search	Free tier at tavily.com
`QDRANT_URL`	Set by Docker Compose	Override for external Qdrant

Tech Stack

Backend — Python 3.11, FastAPI, LangGraph, LangChain, Qdrant, sentence-transformers (multilingual MiniLM + cross-encoder), slowapi, Tavily Frontend — Next.js 15, React 19, Tailwind CSS, TypeScript Infrastructure — Docker Compose (Qdrant + FastAPI + Next.js)

Key Features

Feature	Details
HyDE query rewriting	Before retrieval, the LLM generates a hypothetical answer to the query and retrieves against that — improves recall on short or vague questions
Multilingual embeddings	`paraphrase-multilingual-MiniLM-L12-v2` supports 50+ languages; same 384-dim space, drop-in replacement
Input guardrails	Detects prompt injection / jailbreak patterns, rejects oversized inputs, returns 400 with a clear reason
Output guardrails	Truncates runaway responses, flags hallucination signal phrases
Hybrid retrieval	BM25 + dense vector search combined via `EnsembleRetriever`
Cross-encoder re-ranking	Top-12 candidates re-ranked to top-6 via `ms-marco-MiniLM-L-6-v2`
Session memory	20-turn conversation history per UUID session
Multi-document support	Upload and query multiple PDFs/URLs simultaneously
RAG evaluation	Embedding-based answer relevancy, faithfulness, and context recall scores
Model selector	Switch between OpenAI (gpt-4o-mini/gpt-4o) and HuggingFace (Mistral/Zephyr/Llama-3) at runtime

Chunking Strategies

Strategy	How it works	Best for
Recursive	Splits on `\n\n → \n → . → space`	General documents
Fixed Size	Fixed character count, `separator=" "`	Tabular / code-heavy docs
Semantic	Groups sentences by embedding similarity	Dense academic text
Sentence Window	Sentence + ±2 surrounding sentences	Precise Q&A

RAG Evaluation Metrics

After each completed answer, click Evaluate to get three embedding-based quality scores:

Metric	Measures
Answer Relevancy	How well the answer addresses the question
Faithfulness	How grounded the answer is in the retrieved context
Context Recall	How well the retrieved context covers the question

Scores are computed via cosine similarity between all-MiniLM-L6-v2 embeddings — no external API needed.

License

MIT