Skip to content

MCP query tools: single-word seeding crosses domains, default budget truncates most results #3810

Description

@abderrahim-lectures

graphifyy 0.9.67, graphify.serve MCP, graph of 167k nodes / 889 communities built from a mixed book corpus.

1. Seeding matches single words from the question.
query_graph("incidence algebra derivations poset") seeds on Poset, Incidence, Derivations, Algebra. Incidence resolves to a node from a biostatistics book, so the BFS pulls in an unrelated community. The same happens with "quantum", which matches a quantum machine learning book for a question about quantum Grassmannians. Related to #2384 and #2381, which cover other seed-selection failures; this one is about cross-domain single-token matches. Suggestion: try multi-word phrase matches against labels first, and fall back to single tokens only when no phrase matches.

2. The default token budget drops 85-90% of nodes.
With the default budget, the two queries above showed 35 of 240 and 41 of 400 nodes. The truncation warning is clear, but this is the effect #2530 describes, measured on a 167k-node graph. A larger default (or ranking nodes by relevance to the question before cutting) would help agents that do not retry.

3. shortest_path resolves fuzzy labels to whole-document nodes.
shortest_path("quantum Grassmannian", "incidence algebra") resolved the source to a book-title node. Returning the resolved endpoints, or several candidate matches, would make the result easier to interpret.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions