CFR Compliance Checker

AI-assisted compliance checking system that takes a contract PDF, extracts its clauses, retrieves the relevant CFR (Code of Federal Regulations) sections through a custom eCFR MCP server, and evaluates each clause's compliance using an LLM agent.

Overview

This project has two independent layers:

• MCP Server (src/cfr_compliance_mcp/) — wraps the official eCFR REST API and exposes it as a set of MCP tools (search, section/part/title retrieval, version history, agency lookup).

• Agent Pipeline (agent/) — parses a contract PDF into clauses, builds an optimized CFR search query for each clause, retrieves the matching regulation via the MCP server, and runs an LLM-based compliance evaluation.

The agent layer never talks to eCFR directly — all retrieval goes through the MCP server as an MCP client.

Contract PDF
     │
     ▼
Contract Parser ──▶ Clauses
     │
     ▼
Query Optimizer ──▶ Optimized Query
     │
     ▼
MCP Search (search_regulations → retrieve_section) ──▶ MCP Server ──▶ eCFR REST API
     │
     ▼
Compliance Agent (Agno + Ollama / Llama 3.1)
     │
     ▼
Compliance Result (status, confidence, reason)

Related MCP server: fedreg-mcp

Features

• Custom eCFR MCP server built on FastMCP, with async HTTP client, XML parsing, caching, and typed request/response models.

• Deterministic query optimizer — detects a clause's regulatory domain (environmental, safety, employment, labor, procurement) and predicts likely CFR titles before retrieval, without an LLM call.

• Compliance agent built with Agno running on local Ollama (Llama 3.1), returning a validated Pydantic ComplianceResult (status, confidence, reason).

• End-to-end pipeline that runs a contract PDF through extraction, retrieval, and compliance evaluation with no manual intervention between steps.

MCP Tools

Project Structure

agent/
├── contract_parser.py          # Extracts text from a contract PDF, splits into clauses
├── cfr_query_optimizer.py      # Deterministic domain detection + query optimization
├── mcp_search.py                # MCP client: search_regulations + retrieve_section
├── compliance_agent.py          # Agno/Ollama compliance evaluation, structured output
├── compliance_pipeline.py       # Orchestrates the full clause-by-clause workflow
└── models.py                    # Shared data models (clauses, ComplianceResult, etc.)

src/cfr_compliance_mcp/
├── clients/                     # Async HTTP client for the eCFR API
├── parsing/                     # XML parsing for eCFR API responses
├── tools/                       # MCP tool implementations
└── server.py                    # FastMCP server setup, tool registration, caching

contracts/
└── sample_contract.pdf          # Sample contract for manual/end-to-end testing

tests/
├── test_optimizer_manual.py
└── test_compliance_agent_manual.py

Tech Stack

• Python 3.13

• FastMCP — MCP server framework

• Agno — agent framework with structured (Pydantic) output

• Ollama (Llama 3.1) — local LLM runtime

• Pydantic v2 — validation and typed models

• uv — dependency management and running project modules

Setup

1. Install dependencies:

   uv sync

2. Install and run Ollama locally, and pull the model:

   ollama pull llama3.1

3. Copy .env.example to .env and fill in any required configuration.

Usage

Run the full pipeline against the sample contract:

uv run python -m agent.compliance_pipeline

Run the contract parser / clause extraction step on its own:

uv run python -m agent.workflow

Run the manual test suites:

uv run python -m tests.test_optimizer_manual
uv run python -m tests.test_compliance_agent_manual

Sample Result

For the clause SECTION 2. HAZARDOUS MATERIALS HANDLING, the pipeline correctly matched 40 CFR 262.84 (hazardous waste generator standards) and returned:

Current Capabilities

• Parse a contract PDF and extract its text

• Split contract text into individual, structured clauses

• Deterministically optimize a CFR search query per clause, based on detected regulatory domain

• Search official eCFR content via the custom MCP server

• Retrieve the matched CFR section's full text

• Evaluate clause-vs-regulation compliance using an LLM agent with structured, validated output

• Produce a per-clause compliance result (status, confidence, reason) tied back to the correct clause title

Known Limitations

• No automated CI test suite — testing is currently via manual scripts only

• No multi-clause batch reporting (single clause per run)

• No PDF/Markdown compliance report export

• No ranking mechanism when search_regulations returns multiple plausible candidates — current behavior is best-result selection

• No support yet for clauses governed by more than one CFR section

Roadmap

• Multi-clause contract reports (aggregate results across all clauses)

• PDF/Markdown compliance report export

• Improved retrieval ranking for multiple plausible candidates

• Support for multiple CFR references per clause

• Automated evaluation metrics against a labeled test set

• Automated unit/integration test suite runnable in CI

• REST API / UI for the pipeline

• Normalize LLM confidence values (e.g., convert 95 → 0.95 before validation)

• Detect clauses with no applicable CFR regulation and skip unnecessary retrieval

• Retrieval explainability (optimized query, predicted CFR title, matched keywords, retrieval score)

Branching

Author

Vaibhav Vikas Ranjan