Local-first GraphRAG system exposed as an MCP server
Server has 12 tools with mostly complete schemas and descriptions, but quality is inconsistent. Tool names follow verb_noun convention well (get_, export_, ingest_, query_, list_, submit_, validate_). Descriptions range from adequate (60-150 chars) to verbose (200+ chars). All tools have input schemas with types and descriptions. However, output schemas are not documented, LLMs cannot predict response structure. Error handling is minimal; most tools return generic {status, message} without recovery guidance. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. Parameter descriptions lack format constraints and validation rules. Workspace enum is repeated across 10 tools, suggesting opportunity for consolidation.
Export the local NetworkX graph artifact in GRAPH_IS_HERE to Neo4j as workspace-scoped nodes and relationships.
Export the local vector artifacts in GRAPH_IS_HERE to Qdrant collections for entities, relationships, and chunks.
Return the current MCP runtime status and local graph health summary.
Save a checkpoint of the current ingestion state.
Add a reviewed extraction file to the graph. Ingestion is additive, not replacement.
Additively ingest a reviewed extraction payload (in-memory, no file read)
No output schemas documented. LLMs cannot predict response structure or plan downstream tool calls. All 12 tools return unstructured {status, message} or similar without field documentation.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. export_*, ingest_*, and submit_* are WRITE operations; get_*, list_*, validate_*, query_* are READ_ONLY. Agents cannot distinguish safe from destructive calls.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | <=2025-11-25 | v2 |
Read multiple extraction files, reconcile them into one unified extraction, write unified file to disk, then run the ingestion pipeline.
List entities/relations whose descriptions have outgrown their bounds and need agent compression, with the live content to summarize.
Query the knowledge graph using local retrieval and optional LLM synthesis.
Replay an identical extraction after an operational ingestion failure.
Store the agent's rolling summary and clear (or re-flag) the pending record.
Validate a file with ingestion's non-mutating preflight only.
Credentials exposed as parameters. export_graph_to_neo4j and export_vectors_to_qdrant accept uri, username, password, api_key as tool parameters. These will be logged in agent traces and prompt history. Use server-side secret injection via environment variables.
Error handling lacks recovery guidance. All tools return generic {status: error, message: str} without actionable next steps. E.g., 'file not found' should suggest 'Call ingest_from_file with a valid path' or list available files.
Parameter descriptions lack format constraints. E.g., file_path has no guidance on absolute vs relative paths, supported extensions, or max length. uri, url, api_key lack format hints. Descriptions should state 'absolute path', 'must be .json, .md, or .txt', 'e.g., neo4j+s://localhost:7687'.