ForcePilot/backend/package/yuxi/knowledge/eval/benchmark_generation.py

import asyncio
import json
import random
from collections.abc import AsyncIterator, Callable
from typing import Any

import json_repair

from yuxi.models import select_model
from yuxi.utils import logger

DEFAULT_BENCHMARK_GENERATION_CONCURRENCY = 10
MAX_BENCHMARK_GENERATION_CONCURRENCY = 20


async def collect_kb_chunks(kb_instance: Any, db_id: str) -> list[dict[str, Any]]:
    chunks = []
    for fid, finfo in kb_instance.files_meta.items():
        if finfo.get("database_id") != db_id:
            continue
        try:
            content_info = await kb_instance.get_file_content(db_id, fid)
            for line in content_info.get("lines", []):
                chunks.append(
                    {
                        "id": line.get("id"),
                        "content": line.get("content", ""),
                        "file_id": fid,
                        "chunk_index": line.get("chunk_order_index"),
                    }
                )
        except Exception:
            continue
    return chunks


def clamp_neighbors_count(neighbors_count: int) -> int:
    return min(max(neighbors_count, 0), 10)


def normalize_generation_concurrency_count(value: Any) -> int:
    if value in (None, ""):
        return DEFAULT_BENCHMARK_GENERATION_CONCURRENCY
    return min(max(1, int(value)), MAX_BENCHMARK_GENERATION_CONCURRENCY)


def _is_anchor_chunk(candidate: dict[str, Any], anchor_chunk: dict[str, Any]) -> bool:
    metadata = candidate.get("metadata") or {}
    candidate_id = metadata.get("chunk_id")
    if candidate_id is not None and str(candidate_id) == str(anchor_chunk.get("id")):
        return True

    candidate_file_id = metadata.get("file_id")
    candidate_chunk_index = metadata.get("chunk_index")
    return candidate_file_id == anchor_chunk.get("file_id") and candidate_chunk_index == anchor_chunk.get("chunk_index")


async def select_neighbor_chunks_by_kb_query(
    *, kb_instance: Any, db_id: str, anchor_chunk: dict[str, Any], neighbors_count: int
) -> list[dict[str, Any]]:
    if neighbors_count <= 0:
        return []

    anchor_content = anchor_chunk.get("content", "")
    if not anchor_content:
        return []

    candidates = await kb_instance.aquery(
        anchor_content,
        db_id,
        search_mode="vector",
        final_top_k=neighbors_count + 3,
        use_reranker=False,
        similarity_threshold=0.0,
    )

    chunks = []
    for candidate in candidates:
        if _is_anchor_chunk(candidate, anchor_chunk):
            continue

        metadata = candidate.get("metadata") or {}
        chunk_id = metadata.get("chunk_id")
        content = candidate.get("content", "")
        if not chunk_id or not content:
            continue

        chunks.append(
            {
                "id": str(chunk_id),
                "content": content,
                "file_id": metadata.get("file_id"),
                "chunk_index": metadata.get("chunk_index"),
            }
        )
        if len(chunks) >= neighbors_count:
            break

    return chunks


def build_benchmark_generation_prompt(ctx_items: list[tuple[str, str]]) -> str:
    context_text = "\n\n".join([f"片段ID={cid}\n{content}" for cid, content in ctx_items])
    return (
        "你将基于以下上下文生成一个可由上下文准确回答的问题与标准答案。"
        "仅返回一个JSON对象，不要包含其他文字。"
        "键为 query、gold_answer、gold_chunk_ids。gold_chunk_ids 必须是上述上下文片段的ID子集。\n\n"
        "上下文：\n" + context_text + "\n"
    )


async def _generate_benchmark_item_once(
    *,
    kb_instance: Any,
    db_id: str,
    all_chunks: list[dict[str, Any]],
    llm: Any,
    context_count: int,
) -> dict[str, Any] | None:
    anchor_chunk = all_chunks[random.randrange(len(all_chunks))]
    neighbor_chunks = await select_neighbor_chunks_by_kb_query(
        kb_instance=kb_instance,
        db_id=db_id,
        anchor_chunk=anchor_chunk,
        neighbors_count=context_count - 1,
    )
    ctx_chunks = [anchor_chunk] + neighbor_chunks
    ctx_items = [(chunk["id"], chunk["content"]) for chunk in ctx_chunks]
    allowed_ids = {cid for cid, _ in ctx_items}

    try:
        resp = await llm.call(build_benchmark_generation_prompt(ctx_items), False)
        obj = json_repair.loads(resp.content if resp else "")
        query = obj.get("query")
        answer = obj.get("gold_answer")
        gold_ids = obj.get("gold_chunk_ids")
        if not query or not answer or not isinstance(gold_ids, list):
            logger.warning(f"Generated JSON missing fields or invalid format: {obj}")
            return None

        gold_ids = [str(item) for item in gold_ids if str(item) in allowed_ids]
        if not gold_ids:
            logger.warning("Generated gold_chunk_ids not found in allowed context")
            return None

        return {"query": query, "gold_chunk_ids": gold_ids, "gold_answer": answer}
    except Exception as e:
        logger.warning(f"Benchmark generation failed for one item: {e}")
        return None


async def iter_generated_benchmark_items(
    *,
    kb_instance: Any,
    db_id: str,
    count: int,
    neighbors_count: int,
    llm_model_spec: str | None,
    concurrency_count: int = DEFAULT_BENCHMARK_GENERATION_CONCURRENCY,
    progress_cb: Callable[[int, str], Any] | None = None,
    cancel_cb: Callable[[], Any] | None = None,
) -> AsyncIterator[dict[str, Any]]:
    if progress_cb:
        await progress_cb(5, "加载chunks")

    all_chunks = await collect_kb_chunks(kb_instance, db_id)
    if not all_chunks:
        raise ValueError("No chunks found in knowledge base")

    if progress_cb:
        await progress_cb(15, "准备生成样本")

    if not llm_model_spec:
        raise ValueError("llm_model_spec 不能为空")
    llm = select_model(model_spec=llm_model_spec)
    context_count = max(clamp_neighbors_count(neighbors_count), 1)
    max_attempts = max(count * 5, 50)
    worker_count = normalize_generation_concurrency_count(concurrency_count)
    actual_worker_count = min(worker_count, max(count, 1), max_attempts)
    generated = 0
    results: list[tuple[int, dict[str, Any]]] = []
    state_lock = asyncio.Lock()
    queue: asyncio.Queue[int] = asyncio.Queue()

    for attempt_no in range(max_attempts):
        queue.put_nowait(attempt_no)

    async def worker() -> None:
        nonlocal generated
        while True:
            if cancel_cb:
                await cancel_cb()
            async with state_lock:
                if generated >= count:
                    return
            try:
                attempt_no = queue.get_nowait()
            except asyncio.QueueEmpty:
                return
            try:
                item = await _generate_benchmark_item_once(
                    kb_instance=kb_instance,
                    db_id=db_id,
                    all_chunks=all_chunks,
                    llm=llm,
                    context_count=context_count,
                )
                if item is None:
                    continue
                progress = None
                message = None
                async with state_lock:
                    if generated >= count:
                        continue
                    generated += 1
                    results.append((attempt_no, item))
                    if progress_cb:
                        progress = int(99 * generated / max(count, 1))
                        message = f"已生成 {generated}/{count}"
                if progress_cb:
                    await progress_cb(progress, message)
            finally:
                queue.task_done()

    workers = [asyncio.create_task(worker()) for _ in range(actual_worker_count)]
    try:
        await asyncio.gather(*workers)
    except asyncio.CancelledError:
        for task in workers:
            task.cancel()
        await asyncio.gather(*workers, return_exceptions=True)
        raise
    except Exception:
        for task in workers:
            task.cancel()
        await asyncio.gather(*workers, return_exceptions=True)
        raise

    for _, item in sorted(results, key=lambda pair: pair[0]):
        yield item


def dump_benchmark_item(item: dict[str, Any]) -> str:
    return json.dumps(item, ensure_ascii=False, separators=(",", ":")) + "\n"