#5211 - AI assistant prototype

[reckart]

What's in the PR

• Initial AI assistant implementation accessible on the annotation page as a sidebar
• Basic access to documents in the project via RAG
• Limited access to user manual via RAG
• Updates to the user manual
• Some tests
• Documentation of the early AI assistant
• Refactoring of the Ollama support
• Fix displaying indexing progress on the bulk processing page
• Added re-index button to assistant sidebar
• Normalize embeddings and use dot product instead of cosine when indexing/searching
• Allow suspending background tasks for a project
• During project import and project initialization from a template, suspend background tasks
• During indexing, embed in batches
• Fix timing logging for embeddings
• Allow auto-detecting the embedding dimensions
• Allow using embeddings with a dimension higher than 1024
• Added button to clear conversation
• Aggregate units (sentences) into larger chunks when indexing
• Extract embedding stuff into a separate service and make tokenizer/encoding configurable
• Implement conversation clearing as protocol command
• Factor retrievers out into own classes and introduce an extension point for them
• Introduce a chunker interface
• Add source information to chunks
• UI style changes
• Some classes renameed
• Added time info in assistant messages
• Fix issue when search index has not been initialized when project is being deleted
• Update settings documentation
• Adjust parameters for embedding
• Lay groundworks for source attribution
• Adjust parameters for embedding
• Allow assistant to provide attribution to document context used in responses
• Allow specifying a document when using the DIAM scrollTo function

How to test manually

• Enable the experimental AI assistant
• Make sure Ollama is running and has llama.3.2 and granite-embedding models loaded
• Try talking to it on the annotation page

Automatic testing

• PR includes unit tests

Documentation

• PR updates documentation