v3.3.3: Citations Filtering, Quantized Llama3.2-Vision, GDrive Updates, Title Tooltips, Various UI Enhancements & Bug Fixes

v3.3.3: Citations Filtering, Quantized Llama3.2-Vision, GDrive Updates, Title Tooltips, Various UI Enhancements & Bug Fixes

+ Citations can now be filtered basis the LLM's response by using the appropirate system prompt and checking the 'Filter Citations' Setting in RAG Options!

+ Quantization Support for Llama3.2-Vision multimodal LLMs

• Quanto Float8 quantization level added

• Quant levels displayed in the dropdown are now correctly filtered as per the selected quantization method

• HF-Waitress Auto-Model Detection: Llama3.2-Vision and FLUX.1 auto-detected and necessary flags auto-set, so they user no longer needs to explicitly set model type to Vision or FLUX

• Google Drive Feature Updates:

  1. Search-Filter bar added
  2. Error on authentication token expiration bug fixed and re-login triggers
  3. Logged-in account displayed
  4. Logout button added

• Tables-UI update: GDrive Files Table and Document Loaded to selected VectorDB table have dynamically set min and max height values so as easy UI navigation and enhance look & feel

• Attach file to message for visual analysis by Llama3.2-Vision button now auto-disabled when another LLM is loaded up

• On-hover help Titles added

• Chat name now displayed in Info Box

• Updated requirements.txts

• Various UI refinements

• Various bug fixes: responses no longer overlap, sporadic error on Save Settings changes resolved, etc.

Full Changelog: v3.1...v3.3.3

v3.1: Kosmos MLLM Powered Local AI OCR Comes to LARS-Enterprise!

• Kosmos-2.5 MLLM for fully local AI-OCR finally comes to LARS!! After months of work deploying, containerizing, testing, feature-enriching and stabilizing the Kosmos Multimodal-LLM for it's structured OCR output with no requirements for traditional prompting, Kosmos finally comes to LARS!

Full Changelog: v3.0...v3.1

v3.0: Field Tested & Ready to Deploy!

• Full-fledged Vision-LLM pipeline: utilizing Llama3.2-Vision for VQA, follow-up text-only queries and RAG

• Fullly local AI OCR with Llama3.2-Vision enabling use of one model for document-loading, VQA and RAG in LARS: One Model To Rule Them All!

• Python Waitress WSGI Flask-backend server integrated, enabling queueing of API requests

• BM25F Whoosh-Indexing greatly augmenting RAG pipeline

• Separate Whoosh indexes for every VEctorDB

• Image generation capabilites - support for FLUX.1-Schnell & Dev models added

• UI, UX & QoL Updates: Delete old chats, rename chat titles, filter-serach through chat-history sidebar, new look for Settings fields, etc

• Fixed default configuration and system paths for first-time deployments

• Code improvements - cleanup, reduced imports, modularization and separated, stateless JS

• Successful deployment of LARS-Enterprise with all the above features results in v3.0!

Full Changelog: v2.9.3...v3.0

v2.9.3: New Settings Checkboxes and Revamped Styling

• Added checkbox to enable skipping system prompt

• Added checkbox to force re-extract previously extracted documents for VectorDB loads

• Revamped styling for select dropdowns

• Revamped styling for input text fields

• Changed Settings labels for FLUX-Diffusers for improved clarity

Full Changelog: v2.9.1...v2.9.3

v2.9.1: Filename Bug-Fix and Cleanup

• Added secure_filename function to helper-functions.js, used in createDownloadContainerForFile in chat.js for Vision-LLM chats. This fixs a bug that prevented the download filelink element from working and throwing a File Not Found error since Werkzeug secure_filename is used on the Python backend.

• Un-chaining begins! - Removed unnecessary LangChain imports

Full Changelog: v2.9...v2.9.1

v2.9: Fully Local Vision-LLM backed OCR and Separate Whoosh-BM25F Indexes for Each VectorDB

1. Enabled the use of Llama3.2-Vision LLMs for OCR text-extraction when uploading docs to the VectorDB! Fully local generalised-OCR finally comes to LARS, closing the only loophole to its offline nature, as Azure is no longer necessary for scanned docs!
2. Separate Whoosh-BM25F-Indexes for each VectorDB! Addresses a long desired change, as even the old & now defunct implementation of Whoosh for page-numbers shared one index across all vectorDBs, causing all sorts of pains.
3. Additional code-modularization, code-imporvements, and minor styling updates

Full Changelog: v2.7...v2.9

v2.7: Llama3.2-Vision for RAG & Chats, FLUX Image-Gen, BM25F Indexing, Delete & Rename Chats, Waitress-WSGI Queueing, UI Updates and Code Refactoring

• Llama3.2-Vision for RAG & Chats:

1. Attach files in any format directly to user-input field for analysis via Llama3.2-Vision LLMs
2. Full conversation flow with memory: Ask follow-up questions!
3. RAG-Enabled: Use the same Vision LLM for usual RAG-queries over the exisitng VectorDB+BM25-Index+Re-ranker pipeline!

• FLUX Image-Gen:

1. Fully local image-generation using black-forest-labs/FLUX.1 dev & schnell models
2. Includes options for CPU-offloading and quantization (under 16GB VRAM!)
3. Requires transformers v4.44.0 (below 4.45), modifying hf_waitress.py code to remove MllamaForConditionalGeneration import from transformers and optimum-quanto for quantization

• BM25F Indexing:

1. Whoosh-backed BM25F index create alongside VectorDB chunks when a document is uploaded
2. Both queried at response time and results re-ranked for relevance
3. Pending: separate indexes for each VectorDB, presently, one index is shared by all and clearing one clears the entire index. This limitation will be addressed in a near-future update!

• Code Refactoring:

1. Modularized several functions for improved code-reuse, readibility and maintainability
2. Split JS into 5 separate files for much improved maintainability and readibility
3. Removed various global variables, front-end JS completely stateless

• Delete & Rename Chats in the sidebar

• Waitress-WSGI Queueing to enable multiple, simultaneous chat sessions in different browser tabs/windows

• Various UI & QoL Updates for enhanced look & feel and ease of use