Ingests public consumer-finance complaints (e.g., CFPB CSV).

Uses an LLM to:

Summarize each complaint (2–3 sentences),

Classify into one risk category: Billing, Collections, DataPrivacy, Fees, MisSelling, Fraud, Other,

Extract typed fields (entities, dates, amounts).

Serves a Streamlit app for single and batch analysis.

Tracks latency (and optionally tokens) with MLflow.

finance-ll-complaints/
├─ app/
│  └─ app.py                     # Streamlit UI (OpenAI-only)
├─ src/
│  ├─ ingest.py                  # load/clean raw CFPB CSV → processed CSV
│  ├─ schema.py                  # Pydantic models with validators & normalization
│  ├─ chains.py                  # LangChain pipelines + CLI (single/batch)
│  ├─ evaluate.py                # weak-label evaluation vs Issue/Product
│  └─ utils.py                   # (optional helpers)
├─ data/
│  ├─ raw/                       # raw input CSVs (gitignored)
│  └─ processed/                 # cleaned/scored CSVs (gitignored)
├─ assets/
│  └─ banner.png                 # (optional) project banner
├─ .env.example                  # template for API keys
├─ .gitignore
├─ requirements.txt              # unpinned deps (latest)
└─ README.md    

Python 3.10+ (3.11 recommended)

An OpenAI API key with access to a GPT-4o family model (default: gpt-4o-mini)

OPENAI_API_KEY=sk-...your key...
OPENAI_MODEL=gpt-4o-mini
# Optional MLflow remote (leave blank to use local UI):
MLFLOW_TRACKING_URI=

Use the [CFPB Consumer Complaint Database export (CSV)](https://www.consumerfinance.gov/data-research/consumer-complaints/) Save a file like complaints.csv to:

    data/raw/complaints.csv

Then clean/sample it:
     python src/ingest.py

streamlit run app/app.py

In the app:

Single analysis: paste one complaint → get summary, category, confidence, and extracted fields.

Batch CSV: upload a CSV, choose the text column, pick a row limit, run analysis, preview + download the scored CSV.

MLflow logging: toggle in the sidebar (logs latency; tokens when available).

You can also run the pipeline from the terminal via chains.py:

# Single text
python src/chains.py single -t "Unauthorized charge on my card; bank refused to reverse the fee."

# Batch (limit rows to keep cost/time in check)
python src/chains.py batch -i data/processed/complaints_clean.csv -o data/processed/complaints_scored

Evaluate LLM categories against a keyword-mapped label derived from CFPB Issue + Product:

    python src/evaluate.py

Track latency (and optional tokens) for LLM calls:

    mlflow ui
Open the URL it prints (usually http://127.0.0.1:5000) and look for experiment finance-llm.
To log to a remote server later, set MLFLOW_TRACKING_URI in .env.

Structured outputs: The app uses langchain structured output to coerce LLM responses into a strict Pydantic schema.

Normalization: schema.py normalizes risk_category to one of
Billing | Collections | DataPrivacy | Fees | MisSelling | Fraud | Other.
It also coerces risk_confidence to a float in [0,1] (accepts "90%", "0.9", or 0.9).

Anchoring: The prompt applies precedence rules to reduce drift (e.g., prioritize credit-reporting signals → DataPrivacy).

Add a small few-shot bank of labeled examples to further stabilize categories.

Add cost & token dashboards in the UI.

Swap in vector retrieval for entity normalization (companies, products).

Deploy the Streamlit app to Streamlit Community Cloud or Azure App Service.

(Optional) Expose a FastAPI endpoint for integration.