Advanced retrieval augmented generation for question answering over your own documents, with hybrid retrieval and cross encoder reranking.
Part of the RagFlow line, a series of reference enterprise RAG implementations. This repository is RagFlowPlus, Advanced RAG. See the full line below.
RagFlowPlus lets you upload documents and ask questions about them. It retrieves the relevant passages with hybrid dense and keyword search, reranks them with a cross encoder, sends the best context to a language model, and returns a grounded answer while remembering the conversation. It runs on FastAPI with a Streamlit chat interface and packages the whole stack with Docker.
The clip above is a live, unedited run on a local llama3.2 model with the bundled sample data, including a real SEC 10-K, indexed in Chroma. No paid keys were used. Full recording at assets/videos/ragflowplus-demo.webm, screenshot at assets/screenshots/ragflowplus-ui.png.
The repo ships with sample documents in sample_data, an HR handbook, a product FAQ, and a real SEC 10-K excerpt, so you can run and judge the system without your own files. Fully local and keyless with Ollama:
ollama serve &
ollama pull nomic-embed-text
ollama pull llama3.2:3b
EMBEDDING_PROVIDER=ollama python scripts/load_sample_data.pyThen start the API and UI and ask the questions in sample_data/README.md, including an honesty check where it should decline rather than guess.
| Area | Capability |
|---|---|
| Documents | PDF, DOCX, HTML, TXT, Markdown |
| Embeddings | Google text-embedding-004, or local Ollama nomic-embed-text |
| Retrieval | Dense ChromaDB search plus BM25, fused with Reciprocal Rank Fusion |
| Reranking | Cross encoder ms-marco-MiniLM-L-6-v2 on top of the fused results |
| Memory | Multi turn conversations stored in SQLite |
| Models | OpenAI, Anthropic, or local models via Ollama, chosen by model name |
| Async indexing | Celery worker backed by Redis, so uploads do not block |
| Security | API key auth, rate limiting, HTML input sanitization, CORS |
| Observability | Prometheus metrics at /metrics, structured logging, a health probe |
| Packaging | Docker Compose for the full stack, unit and integration tests, CI |
graph TD
User[User in browser] --> UI[Streamlit chat UI]
UI --> API[FastAPI backend]
API --> Chain[RAG chain]
API --> Memory[(SQLite conversation memory)]
Chain --> Retriever[Hybrid retriever plus reranker]
Retriever --> Chroma[(ChromaDB vector store)]
Chain --> LLM[LLM provider OpenAI, Anthropic, or Ollama]
API --> Worker[Celery worker via Redis]
Worker --> Chroma
cp .env.example .env
# edit .env and set GOOGLE_API_KEY, one LLM key, and a value for API_KEY
docker compose -f docker/docker-compose.yml up --build -d
open http://localhost:8501 # the chat UI
# API docs at http://localhost:8000/docsThen in the UI: upload a document in the sidebar, wait for it to index, and ask a question. The answer comes back grounded in your document, and follow up questions use the earlier conversation.
make install # install dependencies
cp .env.example .env # fill in your keys
docker run -d -p 6379:6379 redis:7-alpine # Redis for the worker
make dev # API at http://localhost:8000
make worker # Celery worker, in a second terminal
make frontend # Streamlit at http://localhost:8501, in a third terminalSettings come from environment variables (see .env.example).
| Variable | Required | Meaning |
|---|---|---|
| GOOGLE_API_KEY | yes | Google AI Studio key for embeddings |
| API_KEY | yes | Sent by clients in the X-API-Key header |
| OPENAI_API_KEY | optional | For GPT models |
| ANTHROPIC_API_KEY | optional | For Claude models |
| OLLAMA_BASE_URL | optional | For local models, default http://localhost:11434 |
| CHROMA_HOST | optional | Set to use a ChromaDB server, otherwise local persistence |
| Method and path | Purpose |
|---|---|
| GET /health | Liveness probe, no auth |
| POST /v1/chat | Ask a question, grounded answer with memory |
| POST /v1/upload-doc | Upload and asynchronously index a document |
| GET /v1/list-docs | List indexed documents |
| POST /v1/delete-doc | Delete a document and its vectors |
| GET /v1/task/{task_id} | Status of an async indexing task |
| GET /metrics | Prometheus metrics |
Python, FastAPI, Streamlit, LangChain, ChromaDB, SQLite, Celery, Redis, Prometheus, and Docker. Embeddings use Google text-embedding-004.
RagFlowPlus is one implementation in the RagFlow line, a series demonstrating distinct enterprise RAG retrieval strategies.
| Year | Repository | Generation |
|---|---|---|
| 2022 | RagFlow | Naive RAG, single dense retrieval |
| 2023 | RagFlowPlus, this repo | Advanced RAG, hybrid retrieval and reranking |
| 2024 | RagFlowPro | Modular production RAG, pgvector, streaming, evaluation |
| 2025 | RagFlowProPlus, RagFlowKAG | Agentic RAG, knowledge graph with reasoning |
| 2026 | RagFlowProMax, UltimateRAG | Multi agent enterprise, multimodal |
The next implementation up, RagFlowPro, replaces ChromaDB with pgvector on Postgres, moves memory to Postgres, computes hybrid retrieval in a single SQL query, streams answers, and adds a measurable evaluation harness.
Every implementation is measured on the same golden questions, keyless, in the rag-catalog hub.
Released under the MIT License. See LICENSE. MIT is the simplest and most permissive of the common licenses, so anyone can read, run, modify, and reuse the code freely.
Malav Patel. GitHub @mlvpatel.