35B 也能逼近万亿模型：小红书 Iris Search Agent 的训练与评测启示

import json
import os
from typing import Any

import requests

LLM_URL = os.environ.get("LLM_URL", "http://localhost:8000/v1/chat/completions")
LLM_MODEL = os.environ.get("LLM_MODEL", "search-agent")
SEARCH_URL = os.environ.get("SEARCH_URL", "http://localhost:9000/search")


def search(query: str, top_k: int = 5) -> list[dict[str, Any]]:
    """假设搜索服务返回 {"results": [{"title", "url", "snippet"}, ...]}。"""
    response = requests.post(
        SEARCH_URL,
        json={"query": query, "top_k": top_k},
        timeout=20,
    )
    response.raise_for_status()
    return response.json().get("results", [])


def ask_llm(messages: list[dict[str, str]]) -> str:
    response = requests.post(
        LLM_URL,
        json={
            "model": LLM_MODEL,
            "temperature": 0.1,
            "messages": messages,
        },
        timeout=60,
    )
    response.raise_for_status()
    return response.json()["choices"][0]["message"]["content"]


def run_agent(question: str, max_rounds: int = 4) -> str:
    evidence: list[dict[str, Any]] = []
    trace: list[dict[str, Any]] = []

    plan = ask_llm([
        {
            "role": "system",
            "content": (
                "你是搜索规划器。把用户问题拆成最多3个可验证的子问题，"
                "只输出JSON数组，例如 [{\"query\": \"...\"}]。"
            ),
        },
        {"role": "user", "content": question},
    ])
    queries = json.loads(plan)

    for round_id, item in enumerate(queries[:max_rounds], start=1):
        query = item["query"]
        results = search(query)
        evidence.extend(results)
        trace.append({"round": round_id, "query": query, "result_count": len(results)})

        decision = ask_llm([
            {
                "role": "system",
                "content": (
                    "你是证据审查器。根据问题和搜索结果判断证据是否足够。"
                    "只输出JSON：{\"enough\": true或false, \"missing\": \"...\"}。"
                ),
            },
            {
                "role": "user",
                "content": json.dumps(
                    {"question": question, "evidence": evidence},
                    ensure_ascii=False,
                ),
            },
        ])
        if json.loads(decision).get("enough"):
            break

    answer = ask_llm([
        {
            "role": "system",
            "content": (
                "根据证据回答问题。不要补写证据中没有的信息；"
                "每个关键结论后标注对应URL。若证据不足，明确说明不确定性。"
            ),
        },
        {
            "role": "user",
            "content": json.dumps(
                {"question": question, "evidence": evidence},
                ensure_ascii=False,
            ),
        },
    ])

    print(json.dumps({"trace": trace, "answer": answer}, ensure_ascii=False, indent=2))
    return answer


if __name__ == "__main__":
    run_agent("比较两个产品在2023年发布的核心功能，并给出一手来源")

pip install requests
export LLM_URL=http://localhost:8000/v1/chat/completions
export LLM_MODEL=your-model
export SEARCH_URL=http://localhost:9000/search
python search_agent.py