OfficeQA Arena Agent

Competition entry for Sentient Arena Challenge 0: OfficeQA.

Architecture

• BM25 Search Index: In-memory inverted index over the full Treasury Bulletin corpus (697 documents). Built at startup, enables fast keyword-based retrieval.
• MCP Server (mcp/treasury/): FastMCP server providing search_corpus, grep_corpus, read_document, search_in_document, list_documents, and document_info tools.
• Prompt Template (prompts/system.j2): Mentor-framed system prompt with structured search/verify/compute protocol, FY/CY rules, and unit handling guidance.
• Skills (skills/): Domain-specific reference guides for search strategy and financial math patterns.
• Harness: Goose agent with MiniMax M2.7 via OpenRouter.

Project Structure

officeqa-arena/
├── arena.yaml              # Agent config
├── mcp/
│   └── treasury/
│       ├── mcp.toml        # MCP server manifest
│       └── server.py       # BM25 search + corpus tools
├── prompts/
│   └── system.j2           # System prompt template
├── skills/
│   ├── treasury-search/
│   │   └── SKILL.md        # Search guide
│   └── financial-math/
│       └── SKILL.md        # Calculation reference
├── pyproject.toml
└── .python-version

Usage

# Authenticate
arena auth login

# Local testing
arena test --smoke

# Submit for evaluation
arena submit