ragchat-ai 0.3.0

RagChat transforms unstructured data for LLM interaction.

RagChat

RagChat enables interaction between large language models and unstructured data. It addresses challenges such as dynamic content updates, varied data sources, and retrieval accuracy by using an upsert-first architecture, filtering mechanisms, and a combination of knowledge graphs with vector search. It supports multi-user, custom models, and self-hosting to provide operational control.

---

Features

• Upsert-first design: Supports constant updates.
• Flexible metadata filtering: Information retrieval allows using custom fields.
• Efficient knowledge graph: Graph is built using small models, promoting efficiency and scalability.
• Multiuser support: Knowledge bases can be isolated or shared.
• Language consistency: Prompts and examples use the specified language, improving reliability.
• Async batch processing: Ingestion of multiple documents can be done in parallel with streaming progress updates.
• Pluggable LLMs and Embedding models: Supports the use of custom models or connection to API endpoints; providers are easily swappable.
• Open source & self-hostable: Operation occurs locally in Docker or directly on a machine, ensuring privacy.

Use Cases

• Casual chat sessions with memories
• Technical documentation search
• Chat+file hybrid RAG with citations
• Personal use
• Multi-user setups

---

Quick start

Docker Compose is required for the easiest setup.

Install RagChat with pip:

pip install ragchat-ai

Recommended local models:

• bge-m3 for embeddings
• qwen3:4b or qwen3:8b for LLM -- Make sure 8k context length is supported.

Configure environment variables:

git clone https://github.com/raul3820/ragchat.git
cd ragchat
cp .env.example .env

Example: Open Webui

Run dependencies with:

docker compose up --build

After startup, the web chat UI is accessible at http://localhost:3001 (refer to .env for port).

Retrieval will be applied to all models. Two flows are presented:

• Casual chat with memories (default)
• Formal RAG with citations (triggered by writing # in the chat and selecting a file)

For file ingestion use: http://localhost:3001/workspace/knowledge

Example: Lihua benchmark with Python SDK

Run dependencies with:

docker compose up neo4j qdrant --build

Once the DB has started, run the file ingestion with:

python -m examples.lihua.step0_index --full

and the Q&A with

python -m examples.lihua.step1_qa --test recall --limit 5

---

Contributing

Contributions welcomed! Please see the CONTRIBUTING.md file for details on how to contribute.

• Bug reports (issues)
• Feature suggestions
• Pull requests (inclusion of tests requested)

Roadmap

Performance:

• Quick retrieval
• Hybrid search
• Multi-hop fact search
• Query intent classification
• Recency weighting
• Better reranking
• Structured aggregates
• 3 phase ingestion (bm25, summaries, fact-entities)
• Graph traversal
• Custom tuning

Flows:

• Chat
• File
• Group chat
• Code
• Web search

Integrations:

• Python SDK
• REST API server
• Neo4j
• Qdrant
• Memgraph? (lower priority)
• Docling
• MCP
• Open-Webui (pipelines)

Testing & Evals:

• LiHua benchmark setup
• LiHua benchmark comparison with other libraries
• Integration test
• Increase test coverage

Security:

• Custom fields sanitization

Documentation:

• Readme/Quick start
• Library documentation
• API documentation

---

Open Source & License

RagChat is MIT-licensed (see LICENSE). Self-hosting and extension are permitted. Certain features may require user-provided LLM/API keys.