Description

Make an autonomous assistant that can respond to user's queries of a collection of BBC News articles based on the type/category of the query.

features:

• automatic and intelligent query routing, based on 4 types of categories:
  • RAG on the articles
  • access to the DB of the articles to answer questions like: how many articles talk about subject A
  • access to various tools for api calls and whatnot
  • general questions based on training date
• TBD

example:

• user: What does the article about Trump tarrifs talk about exactly? does it feature the % of tarrifs on china?
• system: llm routes query to -> rag_query -> queries rag pipeline -> returns answer
• user: ok, can you compare it to tarrifs of 2019?
• system: llm routes query to -> tool_use -> websearch: query = trump china tarrifs 2019

Execution example

---

To run: docker compose run -rm -it assistant and ask your question.

Todo: Add commentMore actions

Assistant:

• Refactor some bad code
• Get Basic rag with routing
  • Custom routing component
• Connect routed query to relevant actor
  • Connect routed query to rag_q
  • Setup agent that accepts tools:
    • Write few tools
    • Connect routed query to tool
    • Connect routed query to db_tool
    • Connect routed query to general qa
    • Connect routed query to tool: web_search

Index:

• Index all data with correct metadata, and different steps

  • Store csv into RDB
  • Store all cells with metadataAdd commentMore actions
  • [-] experiment with different strats (different chunks etc..)

• Retrieval

  • baseline + mrr
    • test data + examples
    • calc mrr
      • write custom mrr component
    • document findings
  • HyDe
  • evaluate
  • automatic metadata labeling + filtering(?)

Setup infra and db

• DB
• Docker/Docker comp

Tests !!! :

• Assistant
• DB
• ETL
• Indexing
  • CSVComponent
  • ...