TinaColab Experiments

This repository collects lightweight Retrieval Augmented Generation (RAG) examples that run against a variety of language models. Each script focuses on a specific step of the workflow—from splitting documents and building embeddings to chatting with remote or local models. When desired, experiments can be tracked with MLflow.

中文說明

Table of Contents

• Requirements
• Setup
• Environment Variables
• Running the Examples
• MLflow Tracking
• Testing

Requirements

• Python 3.10 or newer
• Access to the APIs you plan to call (OpenAI, HuggingFace, Google Gemini, Ollama)
• Optional: an MLflow tracking server or local directory if you wish to log experiments

Setup

1. Install the Python dependencies:

   pip install -r requirements.txt

2. Prepare the data/ directory relative to the repository root:

   • data/med_instruction_v2.md – Markdown instructions used for embedding.
   • data/rfp.pdf – PDF source for rag_example.py.
   • patient_c.txt – optional patient report consumed by the chat examples.

Environment Variables

Configure credentials and optional settings before running the scripts:

• OPENAI_API_KEY – required for OpenAI GPT models.
• HUGGINGFACE_API_KEY – used by 2_1embed_store_query.py and rag_example.py for HuggingFace embeddings (you can also hardcode the value in config.py).
• GEMINI_API_KEY – enables the Google Gemini example.
• MLFLOW_TRACKING_URI (optional) – target MLflow tracking store.

Environment variables can be exported in your shell or stored in a .env file loaded by your preferred tool.

Running the Examples

Each script can be executed independently. Ensure that the data files and vector stores referenced above are in place.

Step	Command	Description
1	python 1_load_split.py	Split data/med_instruction_v2.md into Document chunks.
2	python 2_1embed_store_query.py	Generate embeddings and persist them to ../med_vectordata2/. Run once before querying.
3	python 2_2getvector_query.py	Query the stored vectors from the Chroma database.
4	python 3_1chat_LLM.py	Chat with OpenAI GPT models, optionally providing patient_c.txt as context.
5	python 3_2chat_LLM_local.py	Chat with a local Ollama model at http://localhost:11434/api/chat.
6	python 3_3chat_LLM_gemini.py	Interact with Google Gemini using the embeddings as context.
7	python rag_example.py	Perform full RAG over a PDF, including chunking and local LLM querying.

MLflow Tracking

Launch the MLflow UI to inspect logged runs:

mlflow ui --backend-store-uri ./mlruns

Open the displayed address in your browser to view parameters, metrics and generated responses.

Set MLFLOW_TRACKING_URI to point at a tracking server or directory if you want the scripts to log metrics via the mlflow Python API.

Testing

Offline unit tests cover document ingestion, vector querying and prompt assembly flows. Run them to verify changes:

pytest tests/test_rag_data_scripts.py tests/test_rag_chat_scripts.py

The tests rely on in-memory fakes and mocked API clients, so they do not require network access or real API keys beyond the dummy values supplied by the test harness.