vk_hackathon/search/main.py

import asyncio
import logging
import os
from contextlib import asynccontextmanager
from functools import lru_cache
from typing import Any

import httpx
from fastembed import SparseTextEmbedding
from fastapi import FastAPI, HTTPException, Request
from fastapi.exceptions import RequestValidationError
from fastapi.responses import JSONResponse
from pydantic import BaseModel, Field
from qdrant_client import AsyncQdrantClient, models

EMBEDDINGS_DENSE_MODEL = "Qwen/Qwen3-Embedding-0.6B"

HOST = os.getenv("HOST", "0.0.0.0")
PORT = int(os.getenv("PORT", "8000"))

API_KEY = os.getenv("API_KEY")
EMBEDDINGS_DENSE_URL = os.getenv("EMBEDDINGS_DENSE_URL")
QDRANT_DENSE_VECTOR_NAME = os.getenv("QDRANT_DENSE_VECTOR_NAME", "dense")
QDRANT_SPARSE_VECTOR_NAME = os.getenv("QDRANT_SPARSE_VECTOR_NAME", "sparse")
SPARSE_MODEL_NAME = "Qdrant/bm25"
RERANKER_MODEL = "nvidia/llama-nemotron-rerank-1b-v2"
RERANKER_URL = os.getenv("RERANKER_URL")
OPEN_API_LOGIN = os.getenv("OPEN_API_LOGIN")
OPEN_API_PASSWORD = os.getenv("OPEN_API_PASSWORD")
QDRANT_URL = os.getenv("QDRANT_URL")
QDRANT_COLLECTION_NAME = os.getenv("QDRANT_COLLECTION_NAME", "evaluation")
REQUIRED_ENV_VARS = [
    "EMBEDDINGS_DENSE_URL",
    "RERANKER_URL",
    "QDRANT_URL",
]

logging.basicConfig(level=os.getenv("LOG_LEVEL", "INFO"))
logger = logging.getLogger("search-service")


def validate_required_env() -> None:
    if bool(OPEN_API_LOGIN) != bool(OPEN_API_PASSWORD):
        raise RuntimeError("OPEN_API_LOGIN and OPEN_API_PASSWORD must be set together")

    if not API_KEY and not (OPEN_API_LOGIN and OPEN_API_PASSWORD):
        raise RuntimeError("Either API_KEY or OPEN_API_LOGIN and OPEN_API_PASSWORD must be set")

    missing_env_vars = [
        name for name in REQUIRED_ENV_VARS if os.getenv(name) is None or os.getenv(name) == ""
    ]
    if not missing_env_vars:
        return

    logger.error("Empty required env vars: %s", ", ".join(missing_env_vars))
    raise RuntimeError(f"Empty required env vars: {', '.join(missing_env_vars)}")


validate_required_env()


def get_upstream_request_kwargs() -> dict[str, Any]:
    headers = {"Content-Type": "application/json"}
    kwargs: dict[str, Any] = {"headers": headers}

    if OPEN_API_LOGIN and OPEN_API_PASSWORD:
        kwargs["auth"] = (OPEN_API_LOGIN, OPEN_API_PASSWORD)
        return kwargs

    if API_KEY:
        headers["Authorization"] = f"Bearer {API_KEY}"

    return kwargs


class DateRange(BaseModel):
    from_: str = Field(alias="from")
    to: str


class Entities(BaseModel):
    people: list[str] | None = None
    emails: list[str] | None = None
    documents: list[str] | None = None
    names: list[str] | None = None
    links: list[str] | None = None


class Question(BaseModel):
    text: str
    asker: str = ""
    asked_on: str = ""
    variants: list[str] | None = None
    hyde: list[str] | None = None
    keywords: list[str] | None = None
    entities: Entities | None = None
    date_mentions: list[str] | None = None
    date_range: DateRange | None = None
    search_text: str = ""


class SearchAPIRequest(BaseModel):
    question: Question


class SearchAPIItem(BaseModel):
    message_ids: list[str]


class SearchAPIResponse(BaseModel):
    results: list[SearchAPIItem]


class DenseEmbeddingItem(BaseModel):
    index: int
    embedding: list[float]


class DenseEmbeddingResponse(BaseModel):
    data: list[DenseEmbeddingItem]


class SparseVector(BaseModel):
    indices: list[int] = Field(default_factory=list)
    values: list[float] = Field(default_factory=list)


class ChunkMetadata(BaseModel):
    chat_name: str
    chat_type: str
    chat_id: str
    chat_sn: str
    thread_sn: str | None = None
    message_ids: list[str]
    start: str
    end: str
    participants: list[str] = Field(default_factory=list)
    mentions: list[str] = Field(default_factory=list)
    contains_forward: bool = False
    contains_quote: bool = False


@lru_cache(maxsize=1)
def get_sparse_model() -> SparseTextEmbedding:
    logger.info("Loading local sparse model %s", SPARSE_MODEL_NAME)
    return SparseTextEmbedding(model_name=SPARSE_MODEL_NAME)


@asynccontextmanager
async def lifespan(app: FastAPI):
    app.state.http = httpx.AsyncClient()
    app.state.qdrant = AsyncQdrantClient(
        url=QDRANT_URL,
        api_key=API_KEY,
    )
    try:
        yield
    finally:
        await app.state.http.aclose()
        await app.state.qdrant.close()


app = FastAPI(title="Search Service", version="0.1.0", lifespan=lifespan)

DENSE_PREFETCH_K = 120
SPARSE_PREFETCH_K = 200
RETRIEVE_K = 150
RERANK_LIMIT = 35
KEYWORD_BOOST_EXTRA = 10


async def embed_dense(client: httpx.AsyncClient, text: str) -> list[float]:
    response = await client.post(
        EMBEDDINGS_DENSE_URL,
        **get_upstream_request_kwargs(),
        json={
            "model": os.getenv("EMBEDDINGS_DENSE_MODEL", EMBEDDINGS_DENSE_MODEL),
            "input": [text],
        },
    )
    response.raise_for_status()
    payload = DenseEmbeddingResponse.model_validate(response.json())
    if not payload.data:
        raise ValueError("Dense embedding response is empty")
    return payload.data[0].embedding


async def embed_dense_batch(client: httpx.AsyncClient, texts: list[str]) -> list[list[float]]:
    response = await client.post(
        EMBEDDINGS_DENSE_URL,
        **get_upstream_request_kwargs(),
        json={
            "model": os.getenv("EMBEDDINGS_DENSE_MODEL", EMBEDDINGS_DENSE_MODEL),
            "input": texts,
        },
    )
    response.raise_for_status()
    payload = DenseEmbeddingResponse.model_validate(response.json())
    payload.data.sort(key=lambda x: x.index)
    return [item.embedding for item in payload.data]


def embed_sparse_sync(text: str) -> SparseVector:
    vectors = list(get_sparse_model().embed([text]))
    if not vectors:
        raise ValueError("Sparse embedding response is empty")
    item = vectors[0]
    return SparseVector(
        indices=[int(index) for index in item.indices.tolist()],
        values=[float(value) for value in item.values.tolist()],
    )


def build_dense_query(question: Question) -> str:
    q = question.search_text.strip() if question.search_text else question.text.strip()
    return q


def build_sparse_query(question: Question) -> str:
    base = question.search_text.strip() if question.search_text else question.text.strip()
    parts = [base]
    if question.keywords:
        parts.extend(question.keywords)
    return " ".join(parts)


def _build_keyword_set(question: Question) -> list[str]:
    tokens: list[str] = []
    if question.keywords:
        tokens.extend(kw.lower() for kw in question.keywords if kw)
    if question.entities:
        for field in (
            question.entities.people,
            question.entities.emails,
            question.entities.documents,
            question.entities.names,
            question.entities.links,
        ):
            tokens.extend(e.lower() for e in (field or []) if e)
    return tokens


def prefilter_for_rerank(
    points: list[Any],
    question: Question,
) -> tuple[list[Any], list[Any]]:
    """Select candidates for reranking: top by RRF + keyword-boosted stragglers."""
    if not points:
        return [], []

    head = points[:RERANK_LIMIT]
    tail = points[RERANK_LIMIT:]

    keywords = _build_keyword_set(question)
    if not keywords or not tail:
        return head, tail

    extra: list[Any] = []
    remaining_tail: list[Any] = []
    for p in tail:
        if len(extra) >= KEYWORD_BOOST_EXTRA:
            remaining_tail.append(p)
            continue
        content = ((p.payload or {}).get("page_content") or "").lower()
        if any(kw in content for kw in keywords):
            extra.append(p)
        else:
            remaining_tail.append(p)

    return head + extra, remaining_tail


async def qdrant_search(
    client: AsyncQdrantClient,
    dense_vectors: list[list[float]],
    sparse_vector: SparseVector,
) -> list[Any] | None:
    prefetch_list = []
    for dv in dense_vectors:
        prefetch_list.append(
            models.Prefetch(
                query=dv,
                using=QDRANT_DENSE_VECTOR_NAME,
                limit=DENSE_PREFETCH_K,
            )
        )
    prefetch_list.append(
        models.Prefetch(
            query=models.SparseVector(
                indices=sparse_vector.indices,
                values=sparse_vector.values,
            ),
            using=QDRANT_SPARSE_VECTOR_NAME,
            limit=SPARSE_PREFETCH_K,
        )
    )

    response = await client.query_points(
        collection_name=QDRANT_COLLECTION_NAME,
        prefetch=prefetch_list,
        query=models.FusionQuery(fusion=models.Fusion.RRF),
        limit=RETRIEVE_K,
        with_payload=True,
    )

    if not response.points:
        return None

    return response.points


def extract_message_ids(point: Any) -> list[str]:
    payload = point.payload or {}
    metadata = payload.get("metadata") or {}
    message_ids = metadata.get("message_ids") or []
    return [str(message_id) for message_id in message_ids]


async def get_rerank_scores(
    client: httpx.AsyncClient,
    label: str,
    targets: list[str],
) -> list[float]:
    if not targets:
        return []

    for attempt in range(5):
        try:
            response = await client.post(
                RERANKER_URL,
                **get_upstream_request_kwargs(),
                json={
                    "model": RERANKER_MODEL,
                    "encoding_format": "float",
                    "text_1": label,
                    "text_2": targets,
                },
            )
            if response.status_code == 429:
                wait = 2 ** attempt
                logger.warning(f"Rerank 429, retry {attempt+1}/5 in {wait}s")
                await asyncio.sleep(wait)
                continue
            response.raise_for_status()
            payload = response.json()
            data = payload.get("data") or []
            return [float(sample["score"]) for sample in data]
        except Exception as e:
            logger.warning(f"Rerank error attempt {attempt+1}/5: {e}")
            if attempt < 4:
                await asyncio.sleep(2 ** attempt)
                continue
            logger.error("Rerank failed after 5 attempts, using fallback")
            return []

    logger.error("Rerank 429 after 5 retries, using fallback")
    return []


async def rerank_points(
    client: httpx.AsyncClient,
    query: str,
    points: list[Any],
) -> list[Any]:
    if not points:
        return []
    targets = [point.payload.get("page_content") for point in points]
    scores = await get_rerank_scores(client, query, targets)

    if not scores or len(scores) != len(points):
        logger.warning("Reranker unavailable or score mismatch, returning RRF order")
        return points

    return [
        point
        for _, point in sorted(
            zip(scores, points),
            key=lambda item: item[0],
            reverse=True,
        )
    ]


@app.get("/health")
async def health() -> dict[str, str]:
    return {"status": "ok"}


@app.post("/search", response_model=SearchAPIResponse)
async def search(payload: SearchAPIRequest) -> SearchAPIResponse:
    question = payload.question
    query = question.text.strip()
    if not query:
        raise HTTPException(status_code=400, detail="question.text is required")

    client: httpx.AsyncClient = app.state.http
    qdrant: AsyncQdrantClient = app.state.qdrant

    dense_query = build_dense_query(question)
    sparse_query = build_sparse_query(question)

    dense_task = embed_dense(client, dense_query)
    sparse_task = asyncio.to_thread(lambda: embed_sparse_sync(sparse_query))
    dense_vector, sparse_vector = await asyncio.gather(dense_task, sparse_task)

    dense_vectors = [dense_vector]
    extra_texts: list[str] = []
    raw_text = question.text.strip()
    if raw_text and raw_text != dense_query:
        extra_texts.append(raw_text)
    for v in (question.variants or []):
        q_v = v.strip()
        if q_v and q_v != dense_query and q_v not in extra_texts:
            extra_texts.append(q_v)
    for h in (question.hyde or []):
        q_h = h.strip()
        if q_h and q_h != dense_query and q_h not in extra_texts:
            extra_texts.append(q_h)
    extra_texts = extra_texts[:3]
    if extra_texts:
        try:
            extra_vecs = await embed_dense_batch(client, extra_texts)
            dense_vectors.extend(extra_vecs)
        except Exception as e:
            logger.warning(f"Extra dense embedding failed: {e}")

    all_points = await qdrant_search(qdrant, dense_vectors, sparse_vector)

    if all_points is None:
        return SearchAPIResponse(results=[])

    all_points = list(all_points)

    rerank_pool, rerank_tail = prefilter_for_rerank(all_points, question)
    reranked = await rerank_points(client, query, rerank_pool)
    final_points = reranked + rerank_tail

    msg_score: dict[str, float] = {}
    for rank, point in enumerate(reranked):
        score = 1.0 / (rank + 1)
        for mid in extract_message_ids(point):
            msg_score[mid] = msg_score.get(mid, 0.0) + score
    for rank, point in enumerate(rerank_tail):
        score = 1.0 / (60 + rank + 1)
        for mid in extract_message_ids(point):
            msg_score[mid] = msg_score.get(mid, 0.0) + score
    message_ids = sorted(msg_score, key=lambda m: msg_score[m], reverse=True)[:50]

    return SearchAPIResponse(results=[SearchAPIItem(message_ids=message_ids)])


@app.exception_handler(Exception)
async def exception_handler(request: Request, exc: Exception) -> JSONResponse:
    logger.exception(exc)
    detail = str(exc) or repr(exc)

    if isinstance(exc, RequestValidationError):
        return JSONResponse(status_code=422, content={"detail": exc.errors()})

    if isinstance(exc, HTTPException):
        return JSONResponse(status_code=exc.status_code, content={"detail": exc.detail})

    return JSONResponse(status_code=500, content={"detail": detail})


def main() -> None:
    import uvicorn

    uvicorn.run("main:app", host=HOST, port=PORT, reload=False)


if __name__ == "__main__":
    main()