NexusRAG

A production-oriented corrective, multimodal RAG system, built one tested slice a day.

A 30-day build combining LangGraph orchestration, corrective retrieval with sentence filtering, hybrid dense and sparse search on Qdrant, FastAPI, RAGAS evaluation and LangSmith tracing. Each day adds one tested vertical slice.

_{The project page, with the pipeline, the thirty days and the quick start.}

What runs today

Stage	What it does	Built
Foundation	FastAPI app, typed settings, JSON logging, request IDs, health and readiness endpoints	Days 1 to 3
Ingestion	TXT, Markdown, PDF and DOCX loaders with page citations, sentence-aware chunking	Days 4 and 5
Embeddings	A provider interface, Gemini Embedding 2 and local EmbeddingGemma 2	Days 6 to 8
Semantic search	Indexing and search on Qdrant	Day 9
Keyword search	BM25 on SQLite FTS5	Day 10
Fusion	Reciprocal Rank Fusion over dense and sparse results	Day 11
Reranking	A cross-encoder over the fused candidates	Day 12

Next: LangGraph state, the router agent, corrective grading, generation and grounding. The full plan is in docs/30-day-plan.md.

Planned architecture

Query -> Router -> Hybrid Retrieval -> Correction -> Generation
                                      ^              |
                                      |              v
                              Web fallback <- Reflection -> Grounding

30-day roadmap

The detailed plan is in docs/30-day-plan.md. Days 1 to 12 are complete and published.

Quick start

python -m venv .venv
# Windows: .venv\Scripts\activate
# Linux/macOS: source .venv/bin/activate
pip install -e ".[dev]"
copy .env.example .env       # Windows
# cp .env.example .env       # Linux/macOS
uvicorn app.main:app --reload

Open http://localhost:8000/docs.

Current status

Days 1 to 12 are complete: ingestion, embeddings, semantic and keyword retrieval, RRF fusion, and reranking. Day 12 is documented in docs/day-12-rerank.md.

Decisions are recorded in docs/.

Tests

pip install -e ".[dev,qdrant]"
pytest

The reranker tests inject a fake scorer and download no weights.

Contributing and license

See CONTRIBUTING.md. NexusRAG is released under the MIT License.