bugfix: keep returned results consistent when answering follow-up

[fedhacks]

Description

• feat(server): add new field to CreateMessageReqPayload that accepts a rag_context which overrides the way chunks are currently placed into context window #3927
  • Note: Should not override the chunks JSON that gets returned in response. Also should not override the Clickhouse query. All of that should stay the same.
• feat(search-component): add a tool call to the parallelized set for price and filters that "logs" whether or not this query is a followup
• feat(search-component): add a "buildRAGContext" function which gets called before createMessage to replace the group_ids filter we feed.
• feat(demo-page): add section for configuring followUpTool parameters
• feat(shopify): add options to settings for configuring all of the available tools. Should mirror demo page. #3923

Target(s)

docsearch, shopify

Community channels

Matrix is preferred. Reach out on discord or Matrix for further assistance.

• discord
• matrix

[densumesh]

1. Use tool call w/ the price and tag filter tool call to see if we need to actually search,
2. Add another call to the llm to get a search query generated by the AI
3. do a search w/ the llm generated query and the user query and use the relevance tool to get the best results
4. Modify create_message route to accept a list of chunk_ids and skip searching

[skeptrunedev]

At the moment, if a user asks a follow-up about a set of products, the follow-up generation will search again. This means that the follow-up question's answer will not be actually answering the question!

Results for Answer 1

Results for follow-up question

Let's make it so that we can continue conversations consistently. Please outline your approach when picking this up!