Financial Report Agent

Agent workflow for generating and evaluating financial research reports.

Setup

Requirements:

• Python 3.10+
• wkhtmltopdf if PDF export is needed
• API keys for the configured LLM/VLM providers

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
copy .env.example .env

Edit .env and config.local.yaml as needed. config.local.yaml is preferred over config.yaml when present.

Directories

• data/: runtime memory and reference data.
  • data/memory/long_term/: long-term memory, stock code mapping, and demonstration reports.
  • data/memory/short_term/<model_name>/<stock_code>_<date>/: per-run intermediate materials.
• output/: generated reports and evaluation results.
  • output/reports/<model_name>/: generated .json, .md, and .pdf reports.
  • output/<method_name>_<evaluator_llm_name>_benchmark_results.json: benchmark evaluation outputs.

Reference PDF files are loaded from DEMO_DIR in .env.

Run

Run the single example task in main.py:

python -u main.py

Run benchmark tasks from benchmark.json:

python -u run_benchmark.py --batch_size 1

Evaluate generated benchmark reports:

python -m src.pipelines.evaluation --method_name qwen3-32b

The evaluator reads generated reports from output/reports/<method_name>/ and writes results to output/<method_name>_<evaluator_llm_name>_benchmark_results.json.