Skip to content

ALPHAFLOW_ISSUE_5.md

Latest commit

 

History

History
40 lines (34 loc) · 1.44 KB

ALPHAFLOW_ISSUE_5.md

File metadata and controls

40 lines (34 loc) · 1.44 KB

Mission: Rank the Spirit Fragments (Medium • 4 pts)

The current news_qa_agent.py answers from raw concatenated text. Improve grounding by ranking passages and returning citations.

Goals

  • Chunk input articles into passages (e.g., 120–200 tokens)
  • Rank passages by BM25+ or cosine similarity (tiny embeddings ok)
  • Compose answer from top-K passages
  • Return citations: article source + passage span indices

Scope

  • Pure-Python implementation; no heavy dependencies required
  • Keep API: answer(news_articles, question) returning { answer, citations[] }

Where

  • File: Jujutsu-Quants/app/adk/agents/news_qa_agent.py

Minimal Reproducible Example (MRE)

uvicorn Jujutsu-Quants.app.adk.main:app --reload
curl -s "http://localhost:8000/api/v2/report/simple?question=What%20did%20TSLA%20announce%3F&urls=https://example.com/news1,https://example.com/news2" | jq

Expected: qa.answer summarizes from ranked passages and includes qa.citations like:

{
  "qa": {
    "answer": "Tesla announced ...",
    "citations": [
      {"source": "https://example.com/news1", "start": 350, "end": 520},
      {"source": "https://example.com/news2", "start": 40, "end": 110}
    ]
  }
}

Acceptance Criteria

  • Deterministic chunking and ranking function with tests
  • Answer includes at least one citation when sources exist
  • Graceful fallback when no articles are provided
  • README snippet added in agent section describing behavior