Features · Architecture · Standalone Scripts · Quick Start · Deployment · Structure · Contributing

---

Upload PDFs and ask questions using a Corrective RAG (CRAG) pipeline that evaluates retrieval quality, routes between document and web sources, and synthesizes grounded answers.

Features

• Corrective RAG Pipeline — 6-node LangGraph graph that evaluates each retrieved chunk, then routes between document refinement or web search based on relevance scores
• Multi-Source Routing — CORRECT (score >= 0.7) refines document chunks, INCORRECT (< 0.3) and AMBIGUOUS (mixed) fall back to Tavily web search
• Multi-PDF Ingestion — Upload one or more PDFs, search across all at once
• Sentence-Level Filtering — Refines retrieved chunks into sentences, then keeps only relevant ones before generation
• Dual LLM Fallback Chains — Generation: Mistral -> Groq; Scoring: Gemini with optional Mistral backup for rate limits
• Persistent Vector Store — ChromaDB with HuggingFace embeddings (all-MiniLM-L6-v2)
• Streamlit Chat UI — Clean chat interface with avatars, suggestion chips, and custom styling

Architecture

flowchart LR
    PDF["PDF Upload(s)"] --> Loader["PyPDFLoader"]
    Loader --> Splitter["RecursiveCharacterTextSplitter<br/>(chunk=900, overlap=150)"]
    Splitter --> Embed["HuggingFace Embeddings<br/>(all-MiniLM-L6-v2)"]
    Embed --> Vector[("ChromaDB<br/>Vector Store")]

    Query["User Question"] --> Retrieve["Retrieve<br/>(k=5)"]
    Vector --> Retrieve
    Retrieve --> Eval["Evaluate Each Doc<br/>(scorer LLM)"]

    Eval -->|"score >= 0.7<br/>CORRECT"| Refine["Refine<br/>(filter sentences)"]
    Eval -->|"score < 0.3<br/>INCORRECT"| Rewrite["Rewrite Query"]
    Eval -->|"mixed<br/>AMBIGUOUS"| Rewrite

    Rewrite --> WebSearch["Web Search<br/>(Tavily)"]
    WebSearch --> Generate["Generate Answer<br/>(LLM)"]
    Refine --> Generate

    Generate --> Answer["Answer"]
    Query --> Answer

Loading

Component	Stack
Frontend	Streamlit (chat UI, avatars, suggestion chips)
Orchestration	LangGraph StateGraph (6 nodes, conditional edges)
Document Loading	PyPDFLoader + RecursiveCharacterTextSplitter (900 chars, 150 overlap)
Embeddings	HuggingFace all-MiniLM-L6-v2
Vector Store	ChromaDB (persistent local storage, similarity search k=5)
Generation LLM	Mistral mistral-small-latest -> Groq gpt-oss-120b fallback
Scorer LLM	Gemini gemini-flash-lite-latest (optional Mistral backup via second API key to avoid rate limits)
Web Search	Tavily (lazy-initialized, optional)
Secrets	st.secrets with os.getenv() fallback

Standalone Scripts

app.py and create_database.py are the original single-file versions of the pipeline — the stepping stones that grew into the full ChatPDF app. app.py runs a basic RAG chain using Groq Llama 3.3 70B with MMR retrieval and a simple prompt, no evaluation or web search fallback. create_database.py loads a single PDF, splits it into chunks, and builds the ChromaDB vector store locally.

Use them as a reference for the pipeline logic, or run them standalone:

python create_database.py
python app.py

The full ChatPDF app (frontend.py + backend.py) is the evolved version — a Corrective RAG pipeline with LangGraph, chunk evaluation, web search fallback, sentence-level filtering, and a polished Streamlit chat UI.

Quick Start

git clone https://github.com/kairav7220/ChatPDF.git
cd ChatPDF
pip install -r requirements.txt

Set your API keys in .env:

MISTRAL_API_KEY="your-mistral-key"
MISTRAL_API_KEY_2="your-second-mistral-key"
GROQ_API_KEY="your-groq-key"
GOOGLE_API_KEY="your-google-gemini-key"
HUGGINGFACE_API_TOKEN="your-hf-token"
TAVILY_API_KEY="your-tavily-key"

streamlit run frontend.py

How It Works

1. Upload — Add PDFs via the sidebar and click "Index documents". PDFs are split into chunks (900 chars, 150 overlap) and embedded into ChromaDB.
2. Retrieve — System fetches the 5 most similar chunks from ChromaDB using cosine similarity.
3. Evaluate — Each retrieved chunk is scored for relevance (0.0-1.0) by the scorer LLM. Chunks scoring >= 0.7 are marked as good.
4. Route — Based on scores:
   • CORRECT (at least one chunk >= 0.7): Filter good chunks through sentence-level relevance check, then generate answer from refined context.
   • INCORRECT (all chunks < 0.3): Rewrite the question into a web search query, search via Tavily, generate answer from web results.
   • AMBIGUOUS (mixed scores, none >= 0.7 but not all < 0.3): Same path as INCORRECT — rewrite and search the web.

Deployment

On Streamlit Community Cloud, add your keys in Settings > Secrets:

MISTRAL_API_KEY = "your-mistral-key"
MISTRAL_API_KEY_2 = "your-second-mistral-key"
GROQ_API_KEY = "your-groq-key"
GOOGLE_API_KEY = "your-google-gemini-key"
HUGGINGFACE_API_TOKEN = "your-hf-token"
TAVILY_API_KEY = "your-tavily-key"
CHROMA_DIR = "chroma_db"

Project Structure

ChatPDF/
├── frontend.py              # Streamlit UI (chat, avatars, suggestion chips)
├── backend.py               # CRAG pipeline (LangGraph, 6 nodes, dual LLM chains)
├── app.py                   # Original standalone RAG chain (reference)
├── create_database.py       # Original ChromaDB builder (reference)
├── requirements.txt         # Python dependencies
├── CONTRIBUTING.md          # Contribution guide
├── .env.example             # Required API keys template
├── .gitignore
├── .streamlit/
│   ├── config.toml          # App theme
│   └── secrets.toml         # Your API keys (gitignored)
└── chroma_db/               # Persistent vector store (gitignored)

License

MIT © kairav7220

---

Built with LangChain · LangGraph · ChromaDB · Mistral · Groq · Gemini · Tavily · Sentence Transformers