Added model compatibility for OpenAI and Azure endpoints. Added some Bedrock models, now compatible with thinking models

Changed config defaults to enable successful app run without local model inference on first run

Minor changes to readme repo references and conda environment

Updated pyproject.toml file with requirements from requirements_no_local.txt

Reverted requirements files to previous direct wheel links

Updated git clone reference

Added deduplication with LLM functionality. Minor package updates. Updated installation documentation.

Enhanced user interface and documentation for topic extraction tool. Updated file input labels for clarity, improved README instructions, and refined progress tracking in API calls. Added time tracking for LLM calls in output files.

Added examples for structured summaries and groups. Adapted functions for structured summaries. Simplified front tab GUI

Optimised prompts. Updated Gradio. Added example for zero shot topics. Added support for Granite 4 local model

No longer str capitalize summaries

Corrected blank KV_QUANT_LEVEL values. Removed erroneous extract topics output

Corrected KV quant value when not specified (set to None)

Corrected KV quantisation definition in config. Moved examples to top of screen under intro.

Generally improved inference for low vram systems, unsloth usage improvements, updated packages, switched default local model to Qwen 3 4b

Upload example xlsx files with repo

Updated output logging files. Updated examples and readme. Minor prompt adjustments and package updates

Corrected reasoning setting

Revised intro and readme. Reasoning suffix setting simplified. All in one xlsx returns correct tables.

Position of usage logs input corrected

tracking model state through all_in_one_function

Corrected usage of stop strings, streaming

Added stop strings, optimised llama-cpp-python inference for streaming

Trying out inference with unsloth vs transformers

Added 'all-in-one' function. Corrected local model load when not loaded initially. Environment variables for max data rows and topics.

Added spaces import to llm_funcs

Updated spaces GPU usage for load_model

Added speculative decoding to transformers calls for gemma 3

It should now be possible to load the local model globally at the start to avoid repeated loading throughout the stages of topic extraction

Added framework of support for Azure models (although untested)

Corrected misplaced example file reference. Added AWS Nova models to model list

Minor changes to local inference

Minor corrections to transformers inference

Further optimisations to transformers inference

Minor fixes to llm_funcs

Enabled GPU-based local model inference with the transformers package

Added possibility of adding examples quickly to the input files

Enhanced app functionality by adding new UI elements for summary file management, put in Bedrock model toggle, and refined logging messages. Updated Dockerfile and requirements for better compatibility and added install guide to readme. Removed deprecated code and unnecessary comments.

Corrected backslash

Allowed possibility to run all analysis steps in one click

Updated requirements. More generous spaces GPU timeouts

Removed unnecessary print statements

Loading llama_cpp inside load_model instead of on app load

Enhanced app functionality by adding new logging variables, refining file input options, and updating prompts for better user experience. Updated Dockerfile for improved environment setup and adjusted requirements for compatibility. Removed unnecessary print statements and added error handling in data loading functions.

Removed kv quantisation for CPU as cannot make consistent without flash attention

Aligned kv quantisation for CPU inference

Optimised prompts for llama.cpp prompt caching

Dockerfile now refers to prebuilt llama-cpp-python wheel

Corrected full_prompt save

Minor fixes for Gemini, model calls. Updated Dockerfile for non-GPU systems