80B 로컬 모델 선박에 295 테스트 RAG 코드바를 만들었습니다.

A local-first RAG system built autonomously by a multi-agent framework. This repository is the reference implementation produced by the C.E.H. multi-agent framework — a prompt-based agent cluster (PM, Code, Scaut, Ask, Debug, Writer, Healer) that ships production-grade code with evidence-gated task execution. Every feature below was planned, implemented, tested, and verified by agents following a 13-section task discipline with Definition-of-Done gates. Task files and evidence bundles live in ai_workspace/memory/TASKS/ — the real audit trail, unedited. - Hybrid search — BM25 + dense vectors fused via Reciprocal Rank Fusion (+18.5% accuracy vs vector-only, ~5.9ms latency). See ai_workspace/docs/HYBRID_SEARCH_METRICS.md . - Cross-encoder reranker — cross-encoder/ms-marco-MiniLM-L-6-v2 over top-k results. - Evaluation framework — MRR, NDCG, baseline reports in ai_workspace/evaluation_results/ . - Agentic RAG — self-critique loop with query rewriting. - Tenant isolation — per-tenant filtering, audit logging, Bearer-token auth ( src/security/ ). - Multi-modal — CLIP-based image encoder, unified embedding space, text↔image cross-modal search. - Graph RAG — Neo4j integration with entity extraction and graph traversal ( ai_workspace/docs/GRAPH_RAG.md ). - MCP server — exposes the RAG pipeline to any MCP-compatible client. - LLM: Llama-3-8B-Instruct (Q4_K_M GGUF) via llama-cpp-python - Embeddings: nomic-embed-text-v1.5 (768-dim, multilingual-friendly) - Vector store: ChromaDB / Qdrant (configurable) - Keyword search: BM25 ( rank-bm25 ) - Reranker: sentence-transformers cross-encoder - API: FastAPI with OpenAI-compatible /v1/chat/completions - Framework: LangChain core git clone https://github.com//rag-workshop.git cd rag-workshop/ai_workspace ./install_deps.sh Download the embedding model: python -c "from huggingface_hub import snapshot_download; \ snapshot_download(repo_id='nomic-ai/nomic-embed-text-v1.5', \ local_dir='./models/embeddings', allow_patterns='*.gguf')" Start the llama.cpp servers (embeddings on 8090, LLM on 8080) and run: source .venv/bin/activate python src/mcp_server.py Full setup walkthrough: ai_workspace/INSTRUCTIONS.md . # Unit tests (excludes integration tests marked with @pytest.mark.integration) cd ai_workspace .venv/bin/python -m pytest tests/ # Integration tests (require running llama.cpp + API services) .venv/bin/python -m pytest tests/ -m integration Current state (2026-04-16): 293 passed · 11 failing · 5 skipped out of 309. The 11 failures are tracked as TASK-017, TASK-018 and are being resolved by the C.E.H. agent cluster itself — see the task board for live status. Integration tests (3 tests in test_rag_server.py ) have been marked with @pytest.mark.integration and excluded from default runs via ai_workspace/pytest.ini . rag-workshop/ ├── ai_workspace/ │ ├── src/ │ │ ├── api/ # FastAPI RAG server │ │ ├── agents/ # Agentic RAG components │ │ ├── core/ # Retrievers, rerankers, memory │ │ ├── evaluation/ # MRR / NDCG framework │ │ ├── graph/ # Graph RAG (Neo4j) │ │ ├── multimodal/ # CLIP image pipeline │ │ ├── security/ # Tenant isolation + audit │ │ └── mcp_server.py │ ├── tests/ # 309 tests, ~95% passing │ ├── config/ # YAML configs │ ├── docs/ # Feature deep-dives │ ├── evaluation_results/ # Baseline metrics (evidence) │ ├── memory/ │ │ └── TASKS/ # Every task that built this repo │ └── PROJECT_STATE.md # PM-owned state file ├── README.md # this file └── LICENSE # MIT Each feature corresponds to a numbered task: | Task | What | Status | |---|---|---| | TASK-007 | Hybrid Search (BM25 + vectors, RRF fusion) | DONE | | TASK-008 | Cross-Encoder Reranker | DONE | | TASK-009 | Evaluation Framework (MRR/NDCG) | DONE | | TASK-010 | Agentic RAG patterns | DONE | | TASK-011 | Tenant Isolation + audit logging | DONE | | TASK-012 | Multi-Modal (CLIP) | DONE | | TASK-013 | Graph RAG (Neo4j) | DONE | | TASK-017 | Fix HybridRetriever API mismatch in stress tests | TODO | | TASK-018 | Fix Tenant API integration test route lookup | TODO | | TASK-019 | Mark llama.cpp-dependent tests as integration | TODO | Each task file in ai_workspace/memory/TASKS/ includes the objective, DoD checklist, evidence, and change log. This is what "evidence-gated autonomous development" actually looks like in practice — nothing hidden, nothing polished post-hoc. This repo proves the framework works. If you want the framework itself — the 7 agents, templates, system registry, and custom modes — it's available as a prompt pack: - $29 Starter: full framework, 7 agents, all templates - $49 Pro: Starter + detailed setup + this project as an example - $99/hr Setup Service: I configure C.E.H. for your stack MIT — see LICENSE. Built with C.E.H. — the multi-agent framework that ships code with evidence.