Phase 3: Flow 2 - Recommendation Engine

[jaydee829]

Phase 3: Flow 2 - Recommendation Engine (A2A & MCP)

Goal: Build a user-facing agentic system using A2A for collaboration and MCP for data tools.

• Architecture Strategy:
  • Hybrid Access: Flow 1 (ETL) remains on direct SQLAlchemy for performance. Flow 2 (Agents) uses MCP.
  • Coarse-Grained Tools: MCP tools encapsulate complex logic (e.g., Search + Filter in one call) to ensure ACID compliance and reduce agent-tool chatter.
• Order of Work:
  1. MCP Server Implementation:
     • Expose the Postgres/pgvector database as an MCP server.
     • Implement coarse-grained tools like find_recommendations and log_suggestion.
  2. Experiment: Search Strategies:
     • Implement internal search tool using google-genai.
     • Implement standalone search service with A2A interface.
     • Compare results in MLFlow.
  3. Intelligent Filter Agent:
     • Implement re-read logic and duplicate avoidance via find_recommendations.
  4. Trope-RAG Justification Engine:
     • Build "Justification Prompts" that anchor the LLM's reasoning in retrieved trope facts.
  5. A2A Agent Mesh:
     • Coordinate Search, Filter, Rank agents via A2A messaging.
• Packages: langchain, fastmcp, a2a-sdk (or equivalent LF implementation), google-genai.
• Files/Classes:
  • src/agentic_librarian/mcp/server.py: [NEW] FastMCP server defining tools like get_similar_tropes, check_read_status.
  • src/agentic_librarian/agents/a2a_mesh.py: [NEW] Agent discovery and communication logic following A2A spec.
  • src/agentic_librarian/agents/specialized/: Individual agent logic for search, filter, and rank.