Skip to content

Phase 3 — Agentic capabilities and MCP integration

Status: Planned

Phase 3 extends Alan with richer tool use and exposes it as an MCP server so external AI agents can query the knowledge graph.


Extended tool library

Phase 3 will add higher-level tools that go beyond single-hop retrieval:

Tool (planned) Description
find_path Find the shortest path between two nodes in the graph
expand_neighborhood Return all neighbors of a node up to N hops
aggregate_subgraph Run aggregations over a subgraph matching a pattern
compare_nodes Compare properties and connections of two or more nodes
summarise_cluster Summarise a cluster of related nodes using the LLM

These tools will be added to chat-agent/tools/ following the same pattern as existing tools — each new tool requires only a schema dict and an async function registered in TOOL_MAP.


Multi-step planning

Phase 3 will introduce an optional planner step before the agentic loop: the LLM produces a step-by-step plan for complex questions before executing it. This reduces unnecessary tool calls and improves answer quality on multi-hop questions.


MCP server

Phase 3 will expose Alan as a Model Context Protocol (MCP) server, allowing external AI agents (Claude Desktop, OpenAI Agents, etc.) to use the knowledge graph as a tool.

The MCP server will expose:

Resource / Tool Description
vector_search Semantic similarity search over the indexed graph
graph_query Execute a Cypher query and return results
get_schema Return the graph skeleton as a structured resource
get_node Fetch a single node by ID with its properties and neighbors

The FastAPI server already has an /mcp_tools/ route stub from Phase 1. Phase 3 will complete this implementation using the MCP Python SDK.


Evaluation framework

Phase 3 will add a lightweight evaluation harness to measure retrieval quality and answer accuracy across a curated set of test questions. This makes it possible to compare retrieval strategies (Phase 2 improvements) and tool combinations (Phase 3) objectively.

Metrics:

  • Retrieval recall@k — does the top-k vector search include the ground-truth node?
  • Answer correctness — LLM-as-judge comparison against reference answers
  • Tool call efficiency — how many tool calls does the agent use per question?