Skip to content

README.md

Latest commit

 

History

History
191 lines (135 loc) · 5.64 KB

README.md

File metadata and controls

191 lines (135 loc) · 5.64 KB

🚀 RAG Knowledgebase ChatBot

A Retrieval-Augmented-Generation (RAG) chatbot designed to turn your documents into an intelligent, answerable knowledgebase. Combine vector search over your documents with a powerful language model to provide accurate, context-aware answers — ideal for docs, support knowledgebases, internal wikis, and prototypes.

Short: ingest documents → build vectors → query with a familiar chat interface backed by an LLM.

✨ Highlights

  • Retrieval-Augmented Generation (RAG) architecture: answers grounded in your data, not hallucinated.
  • Pluggable vector store: FAISS, Milvus, Weaviate, Pinecone, etc.
  • Provider-agnostic LLM support: OpenAI, Azure OpenAI, Anthropic, or local LLMs (Llama, Mistral).
  • Document ingestion pipeline: PDFs, Markdown, plain text, HTML, Office docs.
  • Simple API + optional web UI for conversational flows.
  • Tools for chunking, embedding, and context-management (LLM prompt templates, memory window).

📦 What’s inside

  • Ingestion scripts for converting documents to text and creating embeddings.
  • Index builder (vector store) and retrieval layer.
  • Chat service that combines retrieved context with an LLM prompt to generate answers.
  • Example clients (curl / Python) and a minimal web UI (optional) to demo chat flows.
  • Config-driven setup for providers and index backends.

🚀 Quickstart (local)

Prereqs:

  • Python 3.10+ (or your project’s specified runtime)
  • pip or poetry
  • A vector DB or FAISS local index
  • LLM API key (e.g., OpenAI)
  1. Clone
git clone https://github.com/LeelaKarthik-26/RAG_Knowladgebase_ChatBot.git
cd RAG_Knowladgebase_ChatBot
  1. Create virtual env & install
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
  1. Environment variables (example .env)
# LLM provider
LLM_PROVIDER=openai
OPENAI_API_KEY=sk-...

# Vector store choice: faiss | pinecone | weaviate | milvus
VECTOR_STORE=faiss

# If using Pinecone/Weaviate/Milvus supply keys/urls below
PINECONE_API_KEY=
PINECONE_ENV=
WEAVIATE_URL=
MILVUS_HOST=
  1. Ingest your documents
# Example: ingest a folder of docs into the vector store
python scripts/ingest.py --source ./docs --index-name my_knowledgebase
  1. Run the API server
uvicorn app.main:app --reload --port 8000
# or
python -m app.main
  1. Chat with curl
curl -X POST "http://localhost:8000/chat" \
  -H "Content-Type: application/json" \
  -d '{"question":"How do I reset my password?", "index":"my_knowledgebase"}'

🧠 How it works (high level)

  1. Ingest documents → split into chunks with overlap → compute embeddings.
  2. Store embeddings in a vector database.
  3. At query time:
    • Retrieve top-K nearest passages based on embedding similarity.
    • Construct a context-aware prompt (system + retrieved docs + user message).
    • Send prompt to the LLM to generate final answer that cites or references the source passages.

Simple architecture:

  • Ingestion -> Embeddings -> Vector DB
  • API/Chat -> Retriever -> Prompt builder -> LLM -> Response

🔌 Supported components (examples)

  • Embeddings: OpenAI embeddings, local embedding models (e.g., sentence-transformers)
  • Vector stores: FAISS, Pinecone, Weaviate, Milvus
  • LLMs: OpenAI GPT, Azure OpenAI, Anthropic, local LLMs via LLM runtimes
  • File types: .pdf, .md, .txt, .html, .docx

🔧 Configuration & Prompts

  • Prompt templates live in app/prompts/ (or similar); customize to control style, length, and instruction-following.
  • Retrieval settings you’ll likely tweak:
    • top_k (how many passages to retrieve)
    • chunk_size and chunk_overlap during ingestion
    • temperature, max_tokens for LLM generation

Example: Increase top_k to surface more context for complex questions.

✅ Best practices

  • Keep chunks reasonably small (500–1000 tokens) with overlaps of ~100–200 tokens for context continuity.
  • Use citation format or source metadata in the prompt to improve traceability.
  • Limit context length given LLM token limits; use recency or relevance filters if needed.
  • Periodically re-index after document updates.

⚡ Example: Python client

import requests

resp = requests.post(
    "http://localhost:8000/chat",
    json={"question": "What is the refund policy?", "index": "my_knowledgebase"}
)
print(resp.json())

📚 Example prompts

System: "You are a helpful assistant. Use the provided documents to answer the user's question. If the answer cannot be found, say you don't know and provide steps to find the answer."

User + retrieved context: "[RETRIEVED_PASSAGES]\n\nUser: How do I configure SSO for my org?"

🛠️ Development

  • Code style: follow black / isort / flake8 (if present)
  • Tests: run pytest (if tests exist)
  • Add new ingestors under scripts/ or app/ingestors/
  • Add new vector backends by implementing the connector interface in app/vector_store/

🤝 Contributing

Contributions welcome! Please:

  1. Open an issue describing the feature/bug.
  2. Fork and create a feature branch.
  3. Make the change, add tests, and update docs.
  4. Open a pull request referencing the issue.

See CONTRIBUTING.md for more details (if present).

🧾 License

This project is provided under the MIT License. See LICENSE for details.

📬 Need help?

Open an issue or contact the maintainer via GitHub: LeelaKarthik-26

Made with ❤️ for faster, fact-grounded answers from your docs.