Vector search finds an expert that means roughly what you asked for; a graph finds the ones that work together. Running both and scoring on the combination is what lets `expert_find_combinations` propose a team rather than a single best match. The price is real infrastructure — Neo4j alongside the embedded vector store — which is why the Docker path exists. The contexts themselves are your markdown, editable in place and reloaded without a restart.
A registry of expert contexts — markdown knowledge files scoped to a technology — with the machinery to pick the right one automatically. It scans your project to detect what it is built with, matches that against the registry using both vector similarity and graph connectivity, and injects the chosen expert's knowledge into a prompt. Registry and context files live on disk and hot-reload when you edit them.
- `expert_detect_technologies` scans paths in your project to work out what it uses
- `expert_smart_discover` combines vector similarity and graph connectivity to pick the best expert for a task and its constraints
- `expert_semantic_search` and `expert_find_similar` search by meaning rather than by keyword
- `expert_load_context` loads an expert's knowledge and `expert_inject_context` folds it into a prompt
- `expert_explore_network` and `expert_find_combinations` walk expert relationships and propose complementary teams
- `expert_registry_list`, `expert_registry_get` and `expert_registry_search` cover the plain registry lookups
- `expert_track_usage` and `expert_get_analytics` record which expert was used and how it performed
An expert-system directory containing `registry/expert-registry.json`, an `expert-contexts/` folder of markdown, and `performance/metrics.json` — pointed at by `EXPERT_SYSTEM_PATH`. ChromaDB is embedded and creates its own directory; Neo4j is separate and needs `NEO4J_URI` and `NEO4J_PASSWORD`, most easily run in Docker. The README's install line names `expert-registry-mcp` at version 1.0.0, which PyPI does not have; the other documented route is the prebuilt container image, which brings the Neo4j setup and SSE transport with it.