Caseware Procurement Document MCP Server

MCP server over a procurement and inventory knowledge base. Ingests invoices, purchase orders, shipping orders, inventory reports, and contracts, then exposes structured (SQL) and semantic (vector) retrieval via MCP tools for AI agents.

Architecture

Documents              Pipeline                Storage              MCP Server
─────────              ────────                ───────              ──────────
invoices/        ──▶  ingest.py  ──▶  SQLite                search_documents
purchase_orders/ ──▶  extract.py ──▶  ChromaDB  ──▶  stdio ──▶ get_related_documents
shipping_orders/ ──▶  embed.py                      server    find_order_evidence
inventory_reports──▶  index.py                                answer_question
contracts/       ──▶

Hybrid retrieval: structured questions (e.g. "which invoices are missing a PO?") hit SQL; semantic questions (e.g. "summarize contract terms") hit vector search via sentence-transformers, then Ollama (llama3.2) generates grounded answers with citations.

Related MCP server: AXYS MCP Lite

Prerequisites

• Python 3.12+

• Tesseract OCR (for image-based invoice extraction)

  • macOS: brew install tesseract

  • Ubuntu: apt install tesseract-ocr

• Ollama with llama3.2 (for LLM-based answer generation)

  brew install ollama     # macOS
  # or: curl -fsSL https://ollama.com/install.sh | sh
  ollama pull llama3.2

Quick Start

# Clone and enter the project
cd caseware.mcp-codechallege

# Install with pip (no venv required)
pip install -e .

# Run full pipeline + start MCP server
python -m src.caseware_documents_mcp.main

# Or with task (recommended)
task install
task

Dependencies

All listed in pyproject.toml — no requirements.txt needed.

Data

Documents live in data/ organized by type:

data/
├── invoices/             10 PDFs + 4 JPGs
├── purchase_orders/       9 PDFs
├── shipping_orders/      14 PDFs
├── inventory_reports/     7 PDFs
└── contracts/             1 PDF (71 pages)

The pipeline auto-detects format: PDFs via PyMuPDF, images via Tesseract OCR.

Usage

Taskfile (recommended)

Direct

# Full run
python -m src.caseware_documents_mcp.main

# Pipeline only (no server)
python -m src.caseware_documents_mcp.main --pipeline-only

# Skip indexing (if already indexed)
python -m src.caseware_documents_mcp.main --skip-pipeline

MCP Tools

Example questions

• "Which invoices are missing a purchase order?"

• "Find all documents related to order 10687"

• "Summarize the contract terms"

• "Show me all invoices for Roland Mendel"

Claude Desktop Integration

See MCPHOWTO.md for detailed setup instructions.

Project Structure

src/caseware_documents_mcp/
├── models.py            # Pydantic models
├── main.py              # Entry point
├── pipeline/
│   ├── ingest.py        # PDF + OCR extraction
│   ├── extract.py       # Structured field parsing
│   ├── embed.py         # Chunking + embeddings
│   └── index.py         # Cross-document reference builder
├── db/
│   ├── schema.py        # SQLite DDL
│   └── queries.py       # SQL queries
├── retrieval/
│   ├── classifier.py    # Question routing
│   ├── structured.py    # SQL-based retrieval
│   └── semantic.py      # Vector search
└── server/
    └── mcp_server.py    # MCP stdio server

Design Decisions

• SQLite + ChromaDB over a single vector store — procurement data is fundamentally relational; vector search alone misses invoice↔PO↔shipment links

• Regex extraction over LLM-based parsing — the documents follow predictable templates (Northwind-style); regex is faster, cheaper, and more reliable

• sentence-transformers for local embeddings — no API calls needed; all-MiniLM-L6-v2 is 80MB and runs on CPU

• Ollama for answer generation — keeps everything local; llama3.2 provides good results for summarization and reasoning

• Keyword classifier over LLM router — pattern matching on the question is sufficient to distinguish "which invoices are missing a PO?" (SQL) from "summarize the contract" (vector)