Cold Case Investigator

A retrieval-augmented investigator over a folder of case files. Ask a question, it finds the most relevant evidence and writes a report from it, citing the files it used and how closely each one matched.

Repository: https://github.com/sathvik89/Cold_Case_detective

There are two ways to use it: a web UI (app.py) where you can add and remove evidence from the browser, and a terminal version (main.py).

Setup

git clone https://github.com/sathvik89/Cold_Case_detective.git
cd Cold_Case_detective

python3 -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate

pip install -r requirements.txt

Running it

uvicorn app:app --reload

Open http://localhost:8000, click Settings, choose a provider and paste your API key. The key is stored in your browser only.

Terminal version:

cp .env.example .env            # put a key in it first
python main.py

The first question takes a few extra seconds - the embedding model is about 90 MB and downloads once, then it is cached.

Providers

Provider	Get a key	Default model
Groq	https://console.groq.com/keys	llama-3.3-70b-versatile
Google Gemini	https://aistudio.google.com/apikey	gemini-flash-latest
OpenAI	https://platform.openai.com/api-keys	gpt-4o-mini

All three speak the same API format, so switching is just a different base URL and key. The Settings panel can also list the models your key can actually use.

For the terminal version, set one of GROQ_API_KEY, GEMINI_API_KEY or OPENAI_API_KEY in .env, and optionally AI_PROVIDER and AI_MODEL.

Evidence

Evidence is any .txt file in data/. Three sample files ship with the repo. In the web UI you can drop new files in or remove them, and the index rebuilds on the next question. In the terminal version, put files in data/ by hand.

How it works

Each file is embedded with MiniLM and indexed in FAISS using inner product on normalised vectors, which is cosine similarity. A question is embedded the same way, the two closest files are retrieved, and the model writes the report from them. If the evidence does not cover the question it is told to say so rather than guess.

Embedding happens locally, so only the retrieved evidence is ever sent to the provider.

Layout

app.py                  FastAPI server: /api/evidence, /api/ask, /api/providers
providers.py            provider config and client setup
main.py                 terminal version
services/loader.py      reads the .txt files
services/embedder.py    text to vectors
services/retriever.py   FAISS index and search
services/llm.py         prompt building and the provider call
data/                   the evidence files
static/                 the web UI (no build step, plain HTML/CSS/JS)