AI BOOK INTELLIGENCE

Prioritizes chunks found by both BM25 and embeddings, then backfills with reciprocal-rank fusion. The retrieval method remains auditable downstream.

def _intersection_first_hybrid(
    bm25_results: list[SearchResult],
    embed_results: list[SearchResult],
    *,
    top_k: int,
    min_intersection: int = 0,
) -> list[SearchResult]:
    bm25_rank = {r.chunk_id: i for i, r in enumerate(bm25_results)}
    embed_rank = {r.chunk_id: i for i, r in enumerate(embed_results)}
    intersect_ids = [cid for cid in bm25_rank if cid in embed_rank]
    intersect_ids.sort(key=_rrf, reverse=True)

    for cid in intersect_ids:
        base = bm25_by_id.get(cid) or embed_by_id.get(cid)
        out.append(SearchResult(
            chunk_id=cid,
            score=HYBRID_INTERSECT_CONFIDENCE,
            method="hybrid_intersect",
            chapter=base.chapter,
            text=base.text,
        ))

    # Backfill via RRF when intersect is too small.
    for cid, _ in rrf_pool:
        base = bm25_by_id.get(cid) or embed_by_id.get(cid)
        out.append(SearchResult(
            chunk_id=cid,
            score=HYBRID_SINGLE_CONFIDENCE,
            method="hybrid_rrf",
            chapter=base.chapter,
            text=base.text,
        ))

A card cannot validate without source chunks and evidence. Fingerprints, risk flags and confidence travel with the writing material.

class SummaryMaterialCard(SchemaModel):
    card_id: CardIdStr
    query_id: QueryId
    card_type: CardType
    title: str = Field(min_length=1, max_length=80)
    summary: str = Field(min_length=1)

    source_chunks: list[ChunkId] = Field(min_length=1)
    source_fingerprint: FingerprintStr
    evidence: list[EvidenceItem] = Field(min_length=1)
    risk_flags: list[RiskFlag] = Field(default_factory=list)
    confidence: UnitFloat = Field(default=0.0)

Three progressively stricter attempts lower temperature, preserve raw responses for diagnosis and validate normalized cards before acceptance.

retry_configs = [
    {
        "system": "你是学术素材编辑。直接续写 JSON，不要任何前言或解释。",
        "prompt_builder": lambda: _build_llm_prompt(
            branch, query_data, chunk_dicts
        ),
        "temperature": TEMP_MATERIALIZE,
        "max_tokens": max_tokens,
    },
    # Second attempt tightens JSON constraints and lowers temperature.
    # Third attempt uses the minimal fallback prompt at temperature 0.0.
]

for attempt, config in enumerate(retry_configs):
    user_msg, prefill = config["prompt_builder"]()
    response = call_llm(
        client, model, config["system"], user_msg,
        max_tokens=config["max_tokens"],
        temperature=config["temperature"],
        prefill=prefill,
    )
    data = extract_json_from_response(response)
    if data:
        items = _normalize_llm_cards(branch, data, book_slug, qid)
        if items:
            return items