LLMQuant/quant-mind

QuantMind is an agent-native knowledge extraction and retrieval framework for quantitative finance.

What it solves

QuantMind is an information processor designed for quantitative finance. It transforms raw, unstructured financial data—such as research papers, news articles, and filings—into structured, typed knowledge that can be trusted for downstream retrieval and reasoning. It ensures that every piece of extracted knowledge maintains its original citation and timestamp, allowing for precise time-based queries and provenance.

How it works

QuantMind operates through a pipeline of deterministic preprocessing and config-driven operations. It uses a "harness engineering" approach, where the repository itself is designed to be used by coding agents (like Claude or Codex). These agents use repo-level contracts, portable skills, and deterministic verification scripts to build and refine refinement flows.

For developers, it provides a Python library with specific flows:

  • PaperFlow: Converts arXiv PDFs into either a structured hierarchy (structure tree) or a page-aware chunk set with a global summary (semantic shape).
  • collect_news: A tool for gathering news from sources like PR Newswire within specific time windows.
  • batch_run: A utility to execute operations across multiple inputs concurrently.
  • magic: A module to resolve free-form natural language intent into specific flow inputs and configurations.

Who it’s for

  • Quantitative Researchers: Those needing to turn financial papers and news into a structured database for research.
  • AI Agents/Developers: Coding agents and developers who want a reliable, agent-native framework for building financial data pipelines.
  • Financial Data Engineers: Professionals building RAG (Retrieval-Augmented Generation) or agentic RAG systems for finance.

Highlights

  • Agent-Native Design: The repository is built as a "harness" to upgrade general coding agents into domain specialists via contracts and hooks.
  • Source-Faithful Extraction: Uses deterministic preprocessing to ensure exact provenance and replayability without models in the loop.
  • Typed Knowledge Shapes: Produces self-contained artifacts (Paper, News, Earnings, Factor, Thesis) that carry their own metadata.
  • Deterministic Verification: Includes a verify.sh script that enforces linting, types, and tests to ensure code quality before deployment.