enterprise-hybrid-rag-next

Enterprise Hybrid-RAG — Hybrid Retrieval-Augmented Generation (Next Generation)

Status: v1.0.0-pre.1 — production-ready pre-release. All 12 tests pass. ehr = "app.main:main" CLI entry point wired via main(). Duplicate retrieve_async removed.

Part of the We Do Care Global ecosystem — 6 published systems, Apache-2.0 open source, 33% annual corporate revenue pledged to humanitarian assistance (founder-level commitment).

Overview

Enterprise Hybrid-RAG-Next is a development-stage hybrid retrieval-augmented generation project exploring the following architecture and integration areas:

NVIDIA NeMo Retriever — GPU-accelerated dense embeddings via NIM
BM25 + Reciprocal Rank Fusion (RRF) — lexical + semantic hybrid search
Continuous evaluation layer — RAGAS metrics (precision, recall, MRR) with alerts
Multi-modal retrieval — text, charts, structured data in one pipeline
FastAPI + SSE streaming — API and streaming integration area; verify current implementation

Benchmark status: No reproducible performance report is asserted here. Publish the tested commit, dataset, hardware, workload, metric definitions and raw results before quoting precision or latency.

Architecture

enterprise-hybrid-rag-next/
├── app/                              # FastAPI service
│   ├── main.py                       # Entry point, SSE streaming endpoints
│   ├── retriever.py                  # Hybrid retriever (BM25 + NeMo + RRF)
│   ├── models.py                     # Pydantic request/response models
│   ├── config.py                     # Configuration (Qdrant, NIM, etc.)
│   └── __init__.py
├── eval/                             # Evaluation layer
│   ├── ragas_eval.py                 # RAGAS continuous evaluation
│   └── test_rrf_contract.py          # RRF fusion contract tests
├── tests/                            # Unit + integration tests
├── .zenodo.json                      # Zenodo metadata for DOI minting
├── citation.cff                      # Citation file format
├── CHANGELOG.md
├── CODE_OF_CLAUDE.md
├── docker-compose.yml                # Full stack: API + Qdrant + NeMo NIM
├── Dockerfile                        # Multi-stage production image
├── pyproject.toml                    # Python package config
├── requirements.txt
├── requirements.lock
└── .github/workflows/                # CI, Pages, Zenodo, security

Quick Start

Docker Compose (Recommended)

git clone https://github.com/WeDoCareGlobal/enterprise-hybrid-rag-next.git
cd enterprise-hybrid-rag-next
docker-compose up -d
# API at http://localhost:8000
# Health: http://localhost:8000/health

Python Only

pip install -r requirements.txt
python -m app.main
# API at http://localhost:8000

Query the API

curl -X POST http://localhost:8000/query \
  -H "Content-Type: application/json" \
  -d '{"query": "What are the side effects of drug X?", "top_k": 10, "fusion": "rrf"}'

# SSE streaming
curl -N -X POST http://localhost:8000/query/stream \
  -H "Content-Type: application/json" \
  -d '{"query": "explain RAG architecture", "top_k": 5}'

Core Capabilities

Feature	Details
Dense Retrieval	NVIDIA NeMo Retriever embeddings via NIM (nemotron-3-ultra, etc.)
Lexical Retrieval	BM25 with custom tokenization
Fusion	Reciprocal Rank Fusion (RRF) — configurable k
Re-ranking	Cross-encoder reranker (optional)
Vector Store	Qdrant (primary), Milvus (fallback)
Multi-modal	Text + chart + structured data unified indexing
Evaluation	RAGAS: faithfulness, answer_relevancy, context_precision, context_recall
Streaming	SSE for real-time token delivery
Observability	Structured logs, metrics, health endpoints

Configuration

Key environment variables:

# NVIDIA NIM
NIM_API_KEY=...
NIM_EMBEDDING_MODEL=nemotron-3-ultra-550b
NIM_ENABLED=true

# Vector store
QDRANT_URL=http://localhost:6333
QDRANT_COLLECTION=enterprise-rag

# Evaluation
RAGAS_EVAL_INTERVAL=3600  # seconds

# GPU fraction scheduling (NVIDIA NIM)
# GPU fractions allow multiple NIMs to share a single GPU (request = guaranteed,
# limit = burstable). GPU memory swap keeps model weights in CPU when idle,
# delivering 44-61x faster time-to-first-token vs scale-from-zero.
GPU_REQUEST=0.5       # Guaranteed minimum GPU fraction
GPU_LIMIT=1.0         # Burstable GPU fraction limit
GPU_SWAP=true         # Enable GPU memory swap
GPU_SWAP_SIZE=80      # Swap size in GB
GPU_MEMORY_UTILIZATION=0.9  # GPU memory utilization

See app/config.py for full configuration options.

Testing

# Unit tests
pytest tests/ -v

# Integration tests (requires Qdrant + NIM)
pytest tests/ -v -m integration

# RRF contract tests
pytest eval/test_rrf_contract.py -v

# Evaluation
python -m eval.ragas_eval

CI/CD (`.github/workflows/`)

CI (ci.yml): lint (ruff), typecheck (mypy), security (bandit, safety), test with coverage
Security (security.yml): Trivy filesystem scan
Docker (docker.yml): multi-arch build, push to GHCR
Pages (pages.yml): deploy docs/ to GitHub Pages
Zenodo (zenodo.yml): archive releases to Zenodo, mint DOI, push to ORCID

Release & Archival

Every GitHub release is automatically archived to Zenodo (DataCite DOI) and pushed to ORCID (0009-0009-8515-2727).

git tag v1.0.0
git push origin v1.0.0
gh release create v1.0.0 --generate-notes
# → triggers Zenodo workflow → mints DOI → updates ORCID

License

33% of profits → humanitarian aid.

Citation

@software{perla2026hybridrag,
  author = {Perla, Emir},
  title = {Enterprise Hybrid-RAG – Production Retrieval-Augmented Generation with NVIDIA NeMo},
  version = {1.0.0-pre.1},
  year = {2026},
  doi = {10.5281/zenodo.22945802},
  url = {https://github.com/WeDoCareGlobal/enterprise-hybrid-rag-next}
}

WDCG Ecosystem

System	Repo	Purpose
AgentGuard	`WeDoCareGlobal/agentguard-next`	Security & governance control plane
FoundationCentar	`WeDoCareGlobal/FoundationCentar-Next`	One-Click Agent Factory
Enterprise Hybrid-RAG	`WeDoCareGlobal/enterprise-hybrid-rag-next`	This repo — Hybrid retrieval
Agent Eval	`WeDoCareGlobal/agent-eval-next`	Evaluation harness
AgentVault	`WeDoCareGlobal/AgentVault-Next`	Tool registry & assurance
Pharma Intelligence OS	`WeDoCareGlobal/Pharma-Intelligence-Os-Next`	Regulatory intelligence

This site is open source. Improve this page.