ChatVeritas

ChatVeritas is a document-grounded RAG assistant for local knowledge. It uses local retrieval over a private text corpus and only sends generation to the selected backend when needed.

The application separates retrieval and generation:

• Retrieval is always local: question embedding, FAISS search, and cross-encoder reranking.
• Generation can be:
  • a local Hugging Face model with optional PEFT LoRA adapter,
  • LM Studio local OpenAI-compatible generation,
  • or Groq cloud generation.

Key features

• Streamlit multipage UI with Offline Chat, Cloud Chat, and Architecture pages
• Local FAISS retrieval and reranking for grounded answers
• Local model inference with optional PEFT LoRA adapters
• LM Studio local server support for OpenAI-compatible offline generation
• Groq cloud generation for lightweight production experimentation
• Fresh Load sidebar support for clearing cached models and reloading backend
• Utilities for corpus ingestion, chunk inspection, synthetic dataset creation, and QLoRA training
• Centralized settings in configs/settings.py

Quick start

Install

git clone <repository-url>
cd ChatVeritas
python -m pip install -r requirements.txt

Configure

All runtime settings live in configs/settings.py. Use it to adjust:

• data, cache, and model paths
• local and LM Studio generation settings
• Groq model and API key settings
• retrieval, embedding, and reranking parameters
• synthetic question generation and fine-tuning settings

Use .env for secrets. Cloud Chat requires GROQ_API_KEY. LM Studio can use an optional LMSTUDIO_API_KEY when the local server requires auth.

Run

streamlit run app.py

Then use the sidebar to open:

• Offline Chat for Local Model or LM Studio
• Cloud Chat for Groq generation
• Architecture for the built-in architecture reference

Corpus and retrieval

Add UTF-8 .txt files to data/raw/ and build the vector store with:

python scripts/ingest.py

The app uses generated artifacts from data/vectorstore/ at runtime. Re-run ingestion only when the corpus changes.

The architecture reference file data/raw/chatveritas_architecture.txt is part of the searchable corpus. If you update it and want the new content to be retrievable, run python scripts/ingest.py again.

Fresh Load behavior

The Streamlit pages include a sidebar Fresh Load action that clears cached model resources and reloads the selected backend. Use it when you change the model, adapter, LM Studio server, or local runtime state.

Useful commands

# Inspect indexed chunks
python scripts/print_chunks.py

# Build synthetic fine-tuning data
python scripts/prepare_finetune_dataset.py

# Validate fine-tuning dataset
python scripts/fine_tune.py --validate-only

# Train a QLoRA adapter
python scripts/fine_tune.py

Repository structure

ChatVeritas/
|-- app.py
|-- configs/
|   |-- __init__.py
|   '-- settings.py
|-- core/
|   |-- chatveritas.py
|   |-- constants.py
|   |-- exceptions.py
|   |-- llm.py
|   |-- logger.py
|   |-- pipelines.py
|   |-- prompts.py
|   |-- retrieval.py
|   '-- utils.py
|-- pages/
|   |-- 1_Offline_Chat.py
|   |-- 2_Cloud_Chat.py
|   '-- 3_Architecture.py
|-- data/
|   |-- raw/
|   |-- vectorstore/
|   |-- processed/
|   '-- cache/
|-- models/
|   |-- adapters/
|   '-- checkpoints/
|-- scripts/
|-- tests/
'-- requirements.txt

For full architecture and design details, see ChatVeritas_ARCHITECTURE.md.

License

Distributed under the MIT License.