Build tool catalogs
Create stable tool graph artifacts from OpenAPI, MCP, and Python sources.
Official documentation
graph-tool-call Documentation
A technical manual for turning OpenAPI, MCP, and Python tools into contracts, retrieval evidence, target selection, execution plans, quality gates, and trace learning loops.
Manual index
The documentation is organized around the same lifecycle the engine runs in production: build a catalog, search with evidence, select a target, synthesize a plan, validate the result, and learn from traces.
Search and selection
Retrieve compact candidate sets and explain why the final target was selected.
Plan and execute
Turn a selected target into executable steps with structured runtime events.
Validate and learn
Use repeatable gates and scrubbed traces instead of anecdotal quality claims.
Execution model
The library is not another prompt wrapper. Each stage produces an artifact that can be inspected, stored, validated, and passed to an adapter.
01
ToolSchemaIngest
Normalize OpenAPI, MCP, and Python sources into stable tool schemas.
02api_contractContract
Extract request/response fields, auth requirements, enums, and context fields.
03RetrievalResultRetrieve
Rank a small candidate set with keyword, graph, embedding, and annotation signals.
04target_selectorSelect
Compare LLM target output with deterministic action, resource, shape, and contract evidence.
05runner eventsPlan/run
Synthesize executable tool paths, stream structured events, and classify auth/request/API failures.
06suggestionsLearn
Promote scrubbed, validated trace evidence so repeated usage improves future ranking.
Quality gates
Quality work should land with repeatable checks, not intuition. Public claims should link to commands, fixtures, or explicit limitations.
SearchRecall@K, MRR, NDCG, candidate count, Korean/English mixed queriesBenchmark fixtures and `benchmarks/run_benchmark.py` outputOpenAPI buildcontract coverage, semantic coverage, readiness score, stable issue codesOpenAPI readiness report and collection artifact metadataPlan and executeplan hit, runner stages, auth readiness, structured failure reasonsQuality Lab cases with search, plan, and execute modesTrace learningscrubbed payloads, shadow improvement, promotion/rejection statusLearning suggestions and shadow/promotion recordsReleasedocs build, API imports, package version, benchmark claim policyRelease gates before publishing a public claim
Reference paths
Cheat sheetAPI Cheat Sheet
Pick the right API or CLI command for each workflow.
APIPublic APIStable imports and engine-level contracts.
CLICLILocal search, inspect, graph, and diagnostics commands.
SchemasSchemasArtifact, event, and report shapes used by integrations.
LLM contextllms.txtA compact documentation index for LLM-assisted development.