v3.0.5

3.0.5 - 2024-09-19

Fixed

• lock embedding model forward pass (#78) @kyteinsky

v3.0.4

3.0.4 - 2024-09-18

Fixed

• fix remaining lowercase comparisons for COMPUTE_DEVICE @kyteinsky

v3.0.3

3.0.3 - 2024-09-18

Fixed

• use uppercase comparisons for COMPUTE_DEVICE @kyteinsky
• add traceback to caught exception in doc loader @kyteinsky

v3.0.2

3.0.2 - 2024-09-18

Changed

• make stuff fit in 8GB VRAM and don't lock text2text api calls (#70) @kyteinsky

Fixed

• fix: detect additional NVIDIA GPUs (#68) @kyteinsky

v3.0.1

3.0.1 - 2024-08-01

Changed

• update llama-cpp-python package in dockerfile @kyteinsky

Fixed

• nvidia-cuda/llama.cpp compat issue @kyteinsky

v3.0.0

3.0.0 - 2024-07-30

Changed

• New major version to maintain versioning consistency with the companion app
• Update readme @kyteinsky

Added

• Use Taskprocessing TextToText provider as LLM (#60) @marcelklehr
• Upgrade base image to cuda 12.2 @kyteinsky

v2.2.1

2.2.1 - 2024-07-09

Fixed

• use COMPUTE_DEVICE env var if present for config @kyteinsky
• add cuda compat llib path back @kyteinsky

v2.2.0

2.2.0 - 2024-06-25

Fixed

• leave room for generated tokens in the context window @kyteinsky
• Dockerfile llama-cpp-python install @kyteinsky
• Version based repair and other changes (#54) @kyteinsky
• .in.txt and use compiled llama-cpp-python @kyteinsky
• correctly log exceptions @kyteinsky
• do not verify docs before delete in Chroma (#53) @kyteinsky
• offload only when instantiated @kyteinsky
• add odfpy back and update deps @kyteinsky

Changed

• up context limit to 30 @kyteinsky
• update configs @kyteinsky
• change repairs to be version based @kyteinsky
• upgrade base image to cuda 12.1 and drop cuda dev deps @kyteinsky
• gh: run the prompts without strategy matrix @kyteinsky

Added

• simple queueing of prompts @kyteinsky
• dynamic loader and unloader @kyteinsky
• add GET /enabled for init check @kyteinsky
• Use the user's language (#50) @marcelklehr

v2.1.1

2.1.1 - 2024-04-23

Changed

• use 8192 as context length

Fixed

• replace @ with .at. in collection name
• replace pandoc completely due to random memory hogs with other python packages
• types fixes and langchain import updates

v2.1.0

2.1.0 - 2024-04-15

Changed

• no context generation is now a chat completion
• filter sources before document decode

Fixed

• set the memory limit for pandoc to 4GB (nextcloud/context_chat_backend#29)
• adjustments for changes in AppAPI in last two months (nextcloud/context_chat_backend#26)
• pass useContext to the query function
• prune context/query to fit the context window
• pandoc hangs

Added

• accelerator detection on container boot
• repair steps
• increase context length to 16384