v1.1.0

SLM RAG Assistant — v1.1.0

A privacy-first document assistant that runs entirely on your machine. Upload PDF files and ask questions about their contents: the model answers using only what is actually in your documents, with source citations.

What's Changed

• Improved main interface by reducing clutter and moving certain elements to separate modules;
• Created a sliding side panel with additional model settings, such as temperature and chunk size during PDF ingestion;
• Created logging functionality that displays different types of logs in a dedicated logs panel. Also saves to a text file, inside a logs/ folder in the running directory;
• Improved generation script by adding additional query and response quality checks;
• Improved generation instructions for both portuguese and english.

Full Changelog: v1.0.0...v1.1.0

Requirements: Python 3.12+ and a .gguf model file. See the README for setup instructions.

v1.0.0

SLM RAG Assistant — v1.0.0

A privacy-first document assistant that runs entirely on your machine. Upload PDF files and ask questions about their contents: the model answers using only what is actually in your documents, with source citations.

What's included:

• Local inference: uses GGUF language models (Llama, Gemma, DeepSeek, or any compatible model) loaded directly from your computer. No data leaves your machine.
• Retrieval Augmented Generation (RAG): documents are chunked, vectorized, and stored in a local ChromaDB database. A fast query model generates search terms; a reasoning model reads the matched chunks and formulates a grounded answer.
• Two-model pipeline: assign a lightweight model for query generation and a more capable model for reasoning independently.
• PDF document support: select any number of PDFs from your desktop. Documents are ingested once and reused across sessions.
• Document and Financial modes: standard factual extraction, or a mode that allows arithmetic and inference from tabular data (e.g. tax brackets).
• English and Portuguese (BR) support
• Optional HTTP server: expose the assistant on your local network to receive prompts and return answers via JSON, for integration with other applications.
• Session persistence: model paths, selected documents, and settings are restored on next launch.

Requirements: Python 3.12+ and a .gguf model file. See the README for setup instructions.