Local MCP server + data pipeline over a heterogeneous procurement knowledge base
Six well-named, read-only tools with clear, domain-specific descriptions (avg 180 chars). All tools have documented input schemas with types and descriptions. Output schemas are Pydantic models (structured). Naming follows verb_noun pattern (search, get, list, reconcile). However, parameter descriptions lack format constraints (e.g., doc_type enum values are listed in tool description, not schema; order_id lacks range/validation guidance). No error handling guidance visible in tool definitions. No tool annotations (readOnlyHint, idempotentHint). Missing pagination parameters on search/search_contract despite potentially large result sets.
Fetch the full structured record for one document by its id (e.g. 'invoice_10248', 'order_10687', 'batch1-1486'), including line items and a source reference. Use after `search` to drill into a specific hit.
Assemble every document that supports a given order_id — invoice, purchase order and shipping order — with key fields and source refs. Use for "which documents support order 10687?" or "what purchase order supports this invoice?". Reports which document types are missing.
List the available inventory (stock) reports — period, product category and how many products each covers. Use for "what inventory reports are available, and what period do they cover?".
Three-way match across invoice / purchase order / shipping order. Flags missing documents (e.g. invoices with no matching PO) and any quantity/price mismatches between the related records. Pass an order_id for one order, or omit it to reconcile ALL orders. Use for "which invoices are missing a matching PO?", "are there mismatches between invoices, POs and shipping orders?", "does this shipment match the invoice?".
Keyword search for evidence across ALL structured records (invoices, purchase orders, shipping orders, inventory reports, image invoices). Use for "find evidence about a vendor / order / item". Returns ranked hits with short snippets + source refs — call get_document for full detail. Optionally filter by doc_type: invoice | purchase_order | shipping_order | inventory_report | image_invoice | contract.
doc_type parameter in search() lists valid values in description ('invoice | purchase_order | ...') but not as JSON Schema enum. LLMs cannot parse prose constraints reliably.
search() and search_contract() lack pagination guidance. No offset/cursor parameter, no total_count in response. Large result sets risk context window exhaustion.
No tool annotations (readOnlyHint, idempotentHint) in FastMCP registration. All tools are read-only but this is not machine-declared.
order_id parameter in reconcile() and get_order_dossier() lacks validation guidance (range, format). No error message template for invalid IDs.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2025-06-18+ | v2 |
Semantic + keyword (hybrid) search over the supply-of-goods contract. Use for "summarize the contract terms relevant to supply of goods" or any clause-level question. Returns the most relevant clauses with section, page and source ref. Template instruction blocks are excluded from the index; a 'placeholder' highlight means the clause contains unfilled template fields.
No error handling guidance in tool descriptions. If a document_id is not found, what should the LLM do? No recovery hints provided.