Explainable RAG System (DataForge Pipeline)

An explainable Retrieval-Augmented Generation (RAG) pipeline designed to move from "trust the model" to "here's the exact evidence and reasoning trace."

This repo follows the pipeline described in DataForge 2nd Round: DataForge_nikhil24380.pdf.

---

Features

• Local Embeddings — Ollama + Qwen3-embedding:8b for high-quality 4096-dim embeddings (no API calls)
• Domain Classification — Local LLM (Ollama + Qwen2.5) automatically classifies chunks by knowledge domain
• Entity & Relation Extraction — SpaCy transformer model extracts entities and knowledge triples
• Chunking with Metadata — LangChain text splitters with rich metadata preservation
• Vector Store — ChromaDB persistent storage for embeddings
• Observability — Full LangSmith tracing for debugging and transparency
• Explainable Output — Evidence graphs, document contributions, and validation traces
• Fully Local — No external API dependencies for core functionality

---

Quickstart (Windows)

1) Create environment and install dependencies

python -m venv .venv
.venv\Scripts\activate
pip install -U pip
pip install -r requirements.txt

2) Install SpaCy model

python -m spacy download en_core_web_trf

3) Install Ollama and pull the required models

Download Ollama and pull the models:

ollama pull qwen3-embedding:8b
ollama pull qwen2.5:7b

4) Configure environment variables

Create a .env file:

LANGCHAIN_TRACING_V2=true
LANGCHAIN_API_KEY=your_langsmith_api_key
LANGCHAIN_PROJECT=explainable-rag
OLLAMA_HOST=http://localhost:11434
OLLAMA_DOMAIN_MODEL=qwen2.5:7b
OLLAMA_EMBEDDING_MODEL=qwen3-embedding:8b

5) Run the API

uvicorn app.main:app --reload

Access the API docs at http://localhost:8000/docs

---

API Endpoints

Endpoint	Method	Description
/api/v3/ingest/	POST	Upload and process PDF documents
/api/v3/full_pipeline/	POST	Full RAG pipeline (decompose -> retrieve -> graph -> generate)
/api/v3/score_trust/	POST	Calculate trust scores for a generated answer
/api/v3/query_decomposition/	POST	Break a query into sub-queries
/api/v3/extract_entities/	POST	Extract named entities from text
/api/v3/extract_relations/	POST	Extract subject-predicate-object triples
/api/v3/visualize_graph/	POST	Generate interactive knowledge graph visualizations

---

Project Structure

app/
  api/          # Endpoints & dependencies
  core/         # Config & environment settings
  db/           # ChromaDB client
  models/       # Pydantic schemas
  pipeline/     # Stages 1-6 logic
  utils/        # LLM & visualization utilities
  main.py       # FastAPI entry point
data/
  chroma_storage/  # Persistent vector DB
  uploads/         # Uploaded documents
test/            # Test files
requirements.txt

---

Troubleshooting

Ollama Connection Errors

# Check if Ollama is running
ollama list

# Verify models are downloaded
ollama pull qwen3-embedding:8b
ollama pull qwen2.5:7b

SpaCy Model Errors

python -m spacy download en_core_web_trf

---

References

• DataForge_nikhil24380.pdf
• https://docs.langchain.com
• https://arxiv.org/abs/2312.10997

---

License

MIT