Implement AI chat endpoint

[DavidMStraub]

This is a work in progress implementing an LLM-based AI chat assistant endpoint for Gramps Web, based on the RAG (retrieval augmented generation) technique.

Short summary of the architecture:

• gramps_webapi.api.search.text_semantic converts Gramps objects to text for semantic indexing
• v1.0 of Sifts is used for providing a search index for vector embeddings, enabling semantic search
  • The new semantic search index is managed via a --semantic switch on the CLI commands and a semantic boolean query argument on the index endpoints
  • Vector embeddings are computed purely locally via Sentence Transformers
• A new /api/chat/ endpoint
  • Conducts a semantic search on the query
  • Feeds the top 10 search results along with the query to a LLM
  • returns the response

Configuration options:

• VECTOR_EMBEDDING_MODEL: the sentence transformers model to use for vector embeddings. E.g. intfloat/multilingual-e5-small
• LLM_BASE_URL: base URL for openai-python. If empty, uses the OpenAI API; but can use any OpenAI compatible API, e.g. using Ollama locally
• LLM_MODEL: the model to use for chat. E.g., gpt-4o-mini for OpenAI or tinyllama for Ollama.

Guiding principles

• Fully local if desired, optionally leveraging APIs
• Disabled by default (resource consumption, privacy concerns, ...)

What's missing?

• documentation, obviously
• more unit tests
• optional quota system per tree
• limiting chat endpoint to user groups?
• I'm still working on the text_semantic module to complete it for all object types and tweak it to improve semantic search recall and precision, which currently is not as good as I would like
• A config option to limi context length rather than retrieving a fixed number of objects
• Possibly a better way to handle long notes
• Adding a default embedding model to the default docker image

I also plan to

• write a blog post with details about the architecture
• submit a PR to the frontend repo with a chat UI very soon

Note that this is meant to be an initial implementation laying the foundations; this can be made more powerful later on by introducing function calling etc.

[DavidMStraub]

Docs: gramps-project/gramps-web-docs#18

[emyoulation]

This is a work in progress implementing an LLM-based AI chat assistant endpoint for Gramps Web, based on the RAG (retrieval augmented generation) technique.

Please introduce acronyms.
"This is a work in progress implementing a Large Language Model (LLM) based artificial intelligence (AI) chat assistant endpoint for Gramps Web, utilizing the Retrieval-Augmented Generation (RAG) technique."

[DavidMStraub]

Almost done. Expect to finish in the next 7 days.

[DavidMStraub]

• Update API docs for /metadata/

[RahulVadisetty91]

After reviewing gramps_webapi/main.py script

The second definition overwrites the first. It seems the second function is the correct one since it includes the semantic option. Remove the first one to avoid overwriting.