Function _retrieve_report_evidence

agents/reading-agent/main.py:3351–3503

(
    paper: Dict[str, Any],
    user_profile: Dict[str, Any],
    parsed_pdf: Optional[Dict[str, Any]],
)

def _retrieve_report_evidence(
    paper: Dict[str, Any],
    user_profile: Dict[str, Any],
    parsed_pdf: Optional[Dict[str, Any]],
) -> Dict[str, Any]:
    chunks = _build_pdf_chunks(parsed_pdf)
    if not chunks:
        return {}

    profile_summary = _summarize_profile_for_embedding(user_profile)
    title = _clean_text(paper.get("title"))
    abstract = _clean_text(paper.get("abstract"))

    query_specs = [
        {
            "bucket": "background",
            "query": f"{title}\n{abstract}\nresearch background motivation problem challenge prior work",
            "preferred_sections": ("abstract", "introduction", "background"),
        },
        {
            "bucket": "method",
            "query": f"{title}\n{abstract}\ncore method approach model framework algorithm training pipeline",
            "preferred_sections": ("method", "approach", "model", "full_text"),
        },
        {
            "bucket": "results",
            "query": f"{title}\n{abstract}\nmain results experiments evaluation outperform improvement benchmark ablation",
            "preferred_sections": ("results", "experiments", "evaluation"),
        },
        {
            "bucket": "limitations",
            "query": f"{title}\n{abstract}\nlimitations future work discussion challenge failure case",
            "preferred_sections": ("discussion", "limitations", "conclusion"),
        },
        {
            "bucket": "relevance",
            "query": (
                f"{title}\n{abstract}\n"
                f"reader research interests relevance use cases methodology overlap {profile_summary}"
            ),
            "preferred_sections": ("abstract", "introduction", "method", "results"),
        },
    ]

    try:
        embedding_module = _load_embedding_module()
        service = embedding_module.get_embedding_service()
        descriptor = getattr(service, "descriptor", "")
        cache_key = _build_evidence_cache_key(paper, user_profile, parsed_pdf, descriptor)
        cached = _load_cached_retrieved_evidence(cache_key)
        if cached:
            return cached
        query_texts = [item["query"] for item in query_specs]
        chunk_texts = [chunk["text"] for chunk in chunks]
        profile_query_text = profile_summary or title or abstract
        vectors = service.embed_batch(query_texts + chunk_texts + [profile_query_text])
        query_vectors = vectors[: len(query_specs)]
        chunk_vectors = vectors[len(query_specs):len(query_specs) + len(chunk_texts)]
        profile_vector = vectors[-1]