Add RAG (Retrieval Augmented Generation) with citations

[sfloess]

Concept Proven in .claude Testing Ground

Source: claude-global-skills/shared/rag.py

What It Does

Query knowledge base → retrieve relevant chunks → generate answer with citations.

Critical for cross-session learning!

Components

1. Hybrid search (semantic + keyword + RRF)
2. Reranking (bi-encoder → cross-encoder)
3. Context building (from retrieved chunks)
4. Answer generation (LLM with context)
5. Citation tracking (which sources used)

Example

from knowledge_ai import RAG

rag = RAG(collection='my-docs')

# Query
result = rag.query('How does multi-model consensus work?')

print(result.answer)
# "Multi-model consensus uses 3+ AI models..."

print(result.citations)
# [{"source": "consensus-guide.md", "similarity": 0.92}]

Without RAG

• ✅ Store memories
• ✅ Search memories
• ❌ Generate answers FROM memories

With RAG

• ✅ Store memories
• ✅ Search memories
• ✅ Generate answers FROM memories with citations

Benefits

Cross-Session Learning

Session A:

User: "Use arbiter/worker pattern for consensus"
→ Stored with embeddings

Session B:

User: "How should I validate with multiple AIs?"
→ RAG finds Session A memory
→ Responds: "Use arbiter/worker pattern..." [Citation: Session A]

Trust via Citations

• Shows which sources were used
• User can verify answer accuracy
• Transparent retrieval process

Implementation

Add to knowledge-ai:

• rag.py (Python core)
• rag.js (JavaScript for Claude Code)
• Integrate with existing:
  • VectorDB (retrieval)
  • Semantic search (hybrid + reranking)
  • Chunking (context building)
• Add citation tracking
• Add retrieval quality metrics

Architecture

Query → Hybrid Search → Rerank → Build Context → LLM Generate → Citations
         (semantic+keyword)  (precision)  (top-k chunks)   (with context)

Related

• Tested in: claude-global-skills (GitLab)
• Requires: vectordb-ai, semantic-search-ai
• Use case: Cross-session AI learning
• Citations: Build trust and transparency