用户问句
    │
    ▼
① Embedding 召回（bge-m3 + Weaviate）
    按向量相似度取出一批 chunk
    分数高只代表「向量靠近」，不一定是能直接答题的条款
    │
    ▼
② Rerank 精排（bge-reranker-v2-m3）
    把「问句 + 每个 chunk 原文」成对送进交叉编码器
    输出 relevance_score，按分数重排
    只把 Top-K 条交给大模型
    │
    ▼
③ LLM 根据精排后的片段生成答案

"reranking_model": {
  "reranking_provider_name": "langgenius/xinference/xinference",
  "reranking_model_name": "bge-reranker-v2-m3"
}

"reranking_model": {
  "provider": "langgenius/xinference/xinference",
  "model": "bge-reranker-v2-m3"
}

#!/usr/bin/env bash
set -euo pipefail

docker ps --filter name=xinference --format '{{.Names}} {{.Status}} {{.Ports}}'
docker stats xinference --no-stream --format '{{.MemUsage}} CPU={{.CPUPerc}}'
free -h

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 1：确认现状（Rerank 开启前的基线）。"""
from __future__ import annotations

import json
import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


def main() -> int:
    print("=== 1/4 Xinference 容器与内存（在 207 上 ssh 执行 docker 命令）===")
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        for cmd in [
            "docker ps --filter name=%s --format '{{.Names}} {{.Status}} {{.Ports}}'" % C.XF_CONTAINER,
            "docker stats %s --no-stream --format '{{.MemUsage}} CPU={{.CPUPerc}}'" % C.XF_CONTAINER,
            "free -h",
        ]:
            out, err = d.sh(cmd)
            print(out or err)

        print("\n=== 2/4 Xinference 已加载模型（应只有 bge-m3 embedding）===")
        out, err = d.sh(
            "curl -s -H 'Authorization: Bearer %s' %s/v1/models" % (C.XF_KEY, C.XF_URL)
        )
        print(out or err)
        try:
            models = json.loads(out)
            for m in models.get("data", models if isinstance(models, list) else []):
                if isinstance(m, dict):
                    print(" -", m.get("id"), m.get("model_type"), m.get("accelerators"))
        except Exception:
            pass

        print("\n=== 3/4 登录 Dify，读知识库 retrieval_model ===")
        ok, _, csrf = d.login()
        print("login ok:", ok, "csrf:", bool(csrf))
        r = d.get("/datasets/%s" % C.DS_ID)
        pp("dataset retrieval_model_dict", r)
        retr = (r.get("json") or {}).get("retrieval_model_dict") or {}
        if retr.get("reranking_enable"):
            print("[WARN] 知识库已开 Rerank，不是「修改前」基线")

        print("\n=== 4/4 Agent 快照 knowledge.sets[0].retrieval + rerank 模型列表 ===")
        sql = (
            "docker exec docker-db_postgres-1 psql -U postgres -d dify -t -A -c "
            "\"SELECT config_snapshot::jsonb -> 'knowledge' -> 'sets' -> 0 -> 'retrieval' "
            "FROM agent_config_snapshots WHERE id='%s';\"" % C.SNAP_ID
        )
        out, err = d.sh(sql)
        print("agent retrieval:", out.strip() or err)
        r2 = d.get("/workspaces/current/models/model-types/rerank")
        pp("model-types/rerank（期望 data 为空）", r2)
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

{
  "search_method": "semantic_search",
  "reranking_enable": false,
  "reranking_mode": null,
  "reranking_model": {
    "reranking_provider_name": null,
    "reranking_model_name": null
  },
  "top_k": 3,
  "score_threshold_enabled": false,
  "score_threshold": null
}

{
  "mode": "multiple",
  "top_k": 4,
  "reranking_mode": "weighted_score",
  "reranking_enable": false,
  "reranking_model": null
}

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 2：开启 Rerank 前的命中测试（对照基线）。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


QUERY = "法定假日加班的制度是什么"


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        r = d.post(
            "/datasets/%s/hit-testing" % C.DS_ID,
            body={
                "query": QUERY,
                "retrieval_model": {
                    "search_method": "semantic_search",
                    "reranking_enable": False,
                    "top_k": 5,
                    "score_threshold_enabled": False,
                },
            },
        )
        pp("hit-testing before rerank", r)
        records = r.get("json")
        if not isinstance(records, list):
            records = (records or {}).get("records") or []
        print("\n--- 前 5 条 ---")
        for i, rec in enumerate(records[:5], 1):
            seg = (rec.get("segment") or {}).get("content") or ""
            print(i, rec.get("score"), seg[:80].replace("\n", " "))
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env bash
set -euo pipefail

docker exec xinference bash -lc '
set -e
xinference login -e http://127.0.0.1:9997 --username admin --password admin1234
xinference registrations -e http://127.0.0.1:9997 -t rerank
xinference launch -e http://127.0.0.1:9997 \
  -n bge-reranker-v2-m3 \
  -t rerank \
  -u bge-reranker-v2-m3
'

rerank  bge-reranker-v2-m3   ['en', 'zh', 'multilingual']  True
rerank  jina-reranker-v2     ['en', 'zh', 'multilingual']  True

Model uid: bge-reranker-v2-m3

ModelActor(bge-reranker-v2-m3-rep0) loaded
Launch finished: model_name=bge-reranker-v2-m3

#!/usr/bin/env bash
set -euo pipefail

curl -s -H 'Authorization: Bearer xf-K8H3LiS27H4188XdOyWJj74MWMOlIR9tZbiybDbxbSas7eAI' \
  http://127.0.0.1:9997/v1/models

{
  "id": "bge-reranker-v2-m3",
  "model_type": "rerank",
  "model_name": "bge-reranker-v2-m3",
  "model_engine": "sentence_transformers",
  "accelerators": [],
  "language": ["en", "zh", "multilingual"],
  "replica": 1
}

#!/usr/bin/env bash
set -euo pipefail

nvidia-smi
docker inspect xinference --format '{{json .HostConfig.DeviceRequests}} {{.HostConfig.Runtime}}'

#!/usr/bin/env bash
set -euo pipefail

docker exec xinference bash -lc 'nvidia-smi -L'

#!/usr/bin/env bash
set -euo pipefail

docker exec xinference bash -lc '
set -e
xinference login -e http://127.0.0.1:9997 --username admin --password admin1234
xinference registrations -e http://127.0.0.1:9997 -t rerank
xinference launch -e http://127.0.0.1:9997 \
  -n bge-reranker-v2-m3 \
  -t rerank \
  -u bge-reranker-v2-m3 \
  --n-gpu auto
'

#!/usr/bin/env bash
set -euo pipefail

# 仅在已具备「能 launch 的管理员 token」时使用；Dify 用的 xf- Key 不能发这条
curl -s -X POST 'http://127.0.0.1:9997/v1/models' \
  -H 'Authorization: Bearer <管理员或可 launch 的 token>' \
  -H 'Content-Type: application/json' \
  -d '{
    "model_uid": "bge-reranker-v2-m3",
    "model_name": "bge-reranker-v2-m3",
    "model_type": "rerank",
    "n_gpu": "auto"
  }'

#!/usr/bin/env bash
set -euo pipefail

curl -s -H 'Authorization: Bearer xf-K8H3LiS27H4188XdOyWJj74MWMOlIR9tZbiybDbxbSas7eAI' \
  http://127.0.0.1:9997/v1/models

{
  "id": "bge-reranker-v2-m3",
  "model_type": "rerank",
  "model_name": "bge-reranker-v2-m3",
  "model_engine": "sentence_transformers",
  "accelerators": ["0"],
  "replica": 1
}

#!/usr/bin/env bash
set -euo pipefail

docker logs xinference 2>&1 | grep -E 'bge-reranker-v2-m3|cuda|CUDA|Launch finished'
nvidia-smi

#!/usr/bin/env bash
set -euo pipefail

# 在 GPU 主机 116.62.10.183 上
nvidia-smi

#!/usr/bin/env bash
set -euo pipefail

# GPU 版（在 116 上）
docker run -d --name xinference-gpu --restart always --gpus all -p 9997:9997 \
  -e HF_ENDPOINT=https://hf-mirror.com \
  xprobe/xinference:latest

docker exec xinference-gpu bash -lc 'nvidia-smi -L'
docker exec xinference-gpu pip install sentence-transformers
docker logs -f xinference-gpu
# 看到 Uvicorn running on http://0.0.0.0:9997 即就绪

#!/usr/bin/env bash
set -euo pipefail

docker exec xinference-gpu bash -lc '
set -e
xinference login -e http://127.0.0.1:9997 --username admin --password admin1234
xinference launch -e http://127.0.0.1:9997 \
  -n bge-m3 -t embedding -u bge-m3 --n-gpu auto
xinference launch -e http://127.0.0.1:9997 \
  -n bge-reranker-v2-m3 -t rerank -u bge-reranker-v2-m3 --n-gpu auto
'

#!/usr/bin/env bash
set -euo pipefail

curl -s -m 10 http://116.62.10.183:9997/v1/cluster/auth
# 若不通：116 防火墙放行 9997，两机要可路由

#!/usr/bin/env bash
set -euo pipefail

curl -s -H 'Authorization: Bearer <116上的xf-Key>' \
  http://116.62.10.183:9997/v1/models

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 3.2 路径 B：把 Dify 的 Xinference URL 改到 GPU 机（修改后）。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


def patch_credentials(d: Dify, model: str, model_type: str, server_url: str, api_key: str) -> None:
    body = {
        "model": model,
        "model_type": model_type,
        "credentials": {
            "server_url": server_url,
            "model_uid": model,
            "api_key": api_key,
            "invoke_timeout": "60",
            "max_retries": "3",
        },
        "name": model,
    }
    path = "/workspaces/current/model-providers/%s/models/credentials" % C.XINFERENCE_PROVIDER
    r = d.post(path, body)
    pp("POST %s %s" % (model_type, model), r)
    if r.get("code") not in ("200", "201"):
        r = d.req("PUT", path, body=body)
        pp("PUT %s %s" % (model_type, model), r)


def main() -> int:
    gpu_key = C.XF_KEY  # 新起 GPU 容器时换成 116 上生成的 xf- Key
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        print("=== 修改前 URL（207 CPU）===")
        print("embedding/rerank server_url:", C.XF_URL)
        print("\n=== 修改后 URL（116 GPU）===")
        print("embedding/rerank server_url:", C.XF_GPU_URL)
        patch_credentials(d, C.EMBED_MODEL, "text-embedding", C.XF_GPU_URL, gpu_key)
        patch_credentials(d, C.RERANK_MODEL, "rerank", C.XF_GPU_URL, gpu_key)
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env bash
set -euo pipefail

curl -s -m 60 \
  -H 'Authorization: Bearer xf-K8H3LiS27H4188XdOyWJj74MWMOlIR9tZbiybDbxbSas7eAI' \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "bge-reranker-v2-m3",
    "query": "法定假日加班怎么算工资",
    "documents": [
      "法定节假日加班的，加班费 = 固定工资÷应出勤天数×300% ×加班天数。",
      "公司班车发车时间以通知为准。",
      "员工请事假须提前申请。"
    ]
  }' \
  http://127.0.0.1:9997/v1/rerank

{
  "results": [
    {"index": 0, "relevance_score": 0.9980502128601074, "document": null},
    {"index": 1, "relevance_score": 0.00312977796420455, "document": null},
    {"index": 2, "relevance_score": 0.00271617597900331, "document": null}
  ]
}

handle = RESTfulRerankModelHandle(model_uid, server_url, auth_headers)
response = handle.rerank(
    documents=docs,
    query=query,
    top_n=top_n,
    return_documents=True,
)

self.invoke(
    model=model,
    credentials=credentials,
    query="Whose kasumi",
    docs=[
        "Kasumi is a girl's name of Japanese origin meaning \"mist\".",
        "Her music is a kawaii bass, a mix of future bass, pop, and kawaii music ",
        "and she leads a team named PopiParty.",
    ],
    score_threshold=0.8,
)

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 5：把 Rerank 模型加进 Dify 供应商凭据。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        body = {
            "model": C.RERANK_MODEL,
            "model_type": "rerank",
            "credentials": {
                "server_url": C.XF_URL,
                "model_uid": C.RERANK_MODEL,
                "api_key": C.XF_KEY,
                "invoke_timeout": "60",
                "max_retries": "3",
            },
            "name": C.RERANK_MODEL,
        }
        r = d.post(
            "/workspaces/current/model-providers/%s/models/credentials" % C.XINFERENCE_PROVIDER,
            body,
        )
        pp("add rerank cred", r)
        r2 = d.get("/workspaces/current/models/model-types/rerank")
        pp("rerank models", r2)
        return 0 if r.get("code") in ("200", "201") else 1
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

{
  "provider": "langgenius/xinference/xinference",
  "models": [
    {
      "model": "bge-reranker-v2-m3",
      "model_type": "rerank",
      "fetch_from": "customizable-model",
      "status": "active"
    }
  ]
}

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 6 · 修改前：只读知识库 retrieval_model（不要 PATCH）。"""
from __future__ import annotations

import json
import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        r = d.get("/datasets/%s" % C.DS_ID)
        pp("dataset before patch", r)
        retr = (r.get("json") or {}).get("retrieval_model_dict") or {}
        print("\n期望修改前关键字段：")
        print(json.dumps({
            "reranking_enable": retr.get("reranking_enable"),
            "top_k": retr.get("top_k"),
            "reranking_mode": retr.get("reranking_mode"),
            "reranking_model": retr.get("reranking_model"),
        }, ensure_ascii=False, indent=2))
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 6 · 修改后：PATCH 知识库，开启 Rerank，Top-K=5。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


def build_retrieval_after() -> dict:
    return {
        "search_method": "semantic_search",
        "reranking_enable": True,
        "reranking_mode": "reranking_model",
        "reranking_model": {
            "reranking_provider_name": C.XINFERENCE_PROVIDER,
            "reranking_model_name": C.RERANK_MODEL,
        },
        "top_k": 5,
        "score_threshold_enabled": False,
        "score_threshold": None,
    }


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        ds_retrieval = build_retrieval_after()
        r = d.req("PATCH", "/datasets/%s" % C.DS_ID, body={"retrieval_model": ds_retrieval})
        print("HTTP", r.get("code"))
        j = r.get("json") or {}
        print("retrieval_model_dict:", j.get("retrieval_model_dict"))
        pp("patch dataset", r)
        return 0 if r.get("code") == "200" else 1
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

{
  "search_method": "semantic_search",
  "reranking_enable": true,
  "reranking_mode": "reranking_model",
  "reranking_model": {
    "reranking_provider_name": "langgenius/xinference/xinference",
    "reranking_model_name": "bge-reranker-v2-m3"
  },
  "top_k": 5,
  "score_threshold_enabled": false,
  "score_threshold": null
}

{
  "mode": "multiple",
  "model": null,
  "top_k": 5,
  "weights": {
    "weight_type": "customized",
    "vector_setting": {
      "vector_weight": 1,
      "embedding_model_name": "bge-m3",
      "embedding_provider_name": "langgenius/xinference/xinference"
    },
    "keyword_setting": {
      "keyword_weight": 0
    }
  },
  "reranking_mode": "reranking_model",
  "reranking_model": {
    "provider": "langgenius/xinference/xinference",
    "model": "bge-reranker-v2-m3"
  },
  "score_threshold": null,
  "reranking_enable": true
}

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 7 · 修改前：备份 Agent 快照/草稿 retrieval（只读）。"""
from __future__ import annotations

import json
import sys
from pathlib import Path

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify


def load_cfg(d: Dify, table: str, cid: str) -> dict:
    out, err = d.sh(
        "docker exec docker-db_postgres-1 psql -U postgres -d dify -t -A -c "
        "\"SELECT config_snapshot FROM %s WHERE id='%s'\"" % (table, cid)
    )
    raw = out.strip()
    if not raw:
        raise RuntimeError(err or "empty config_snapshot")
    return json.loads(raw)


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        snap = load_cfg(d, "agent_config_snapshots", C.SNAP_ID)
        draft = load_cfg(d, "agent_config_drafts", C.DRAFT_ID)
        bak_dir = Path(__file__).resolve().parent / "backup"
        bak_dir.mkdir(exist_ok=True)
        for name, cfg in [("snap_before", snap), ("draft_before", draft)]:
            p = bak_dir / ("专题十三_07_%s.json" % name)
            p.write_text(json.dumps(cfg, ensure_ascii=False, indent=2), encoding="utf-8")
            retr = cfg.get("knowledge", {}).get("sets", [{}])[0].get("retrieval", {})
            print(name, "retrieval:", json.dumps(retr, ensure_ascii=False))
            print("  -> 已备份", p)
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 7 · 修改后：同步改 Agent 快照与草稿的 knowledge.sets[0].retrieval。"""
from __future__ import annotations

import json
import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify


def load_cfg(d: Dify, table: str, cid: str) -> dict:
    out, err = d.sh(
        "docker exec docker-db_postgres-1 psql -U postgres -d dify -t -A -c "
        "\"SELECT config_snapshot FROM %s WHERE id='%s'\"" % (table, cid)
    )
    raw = out.strip()
    if not raw:
        raise RuntimeError(err or "empty config_snapshot")
    return json.loads(raw)


def save_cfg(d: Dify, table: str, cid: str, cfg: dict, tag: str) -> None:
    sql = "UPDATE %s SET config_snapshot = $cfg$%s$cfg$ WHERE id='%s';" % (
        table,
        json.dumps(cfg, ensure_ascii=False),
        cid,
    )
    remote = "/tmp/upd_%s.sql" % tag
    with d.sftp.open(remote, "w") as f:
        f.write(sql)
    out, err = d.sh("docker exec -i docker-db_postgres-1 psql -U postgres -d dify < %s" % remote)
    print(table, (out or err).strip())


def enable_agent_rerank(cfg: dict) -> dict:
    retr = cfg["knowledge"]["sets"][0].setdefault("retrieval", {})
    retr["top_k"] = 5
    retr["reranking_enable"] = True
    retr["reranking_mode"] = "reranking_model"
    retr["reranking_model"] = {
        "provider": C.XINFERENCE_PROVIDER,
        "model": C.RERANK_MODEL,
    }
    return cfg


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        snap = load_cfg(d, "agent_config_snapshots", C.SNAP_ID)
        draft = load_cfg(d, "agent_config_drafts", C.DRAFT_ID)
        save_cfg(d, "agent_config_snapshots", C.SNAP_ID, enable_agent_rerank(snap), "snap_rerank")
        save_cfg(d, "agent_config_drafts", C.DRAFT_ID, enable_agent_rerank(draft), "draft_rerank")
        out, _ = d.sh(
            "docker exec docker-db_postgres-1 psql -U postgres -d dify -t -A -c "
            "\"SELECT config_snapshot::json -> 'knowledge' -> 'sets' -> 0 -> 'retrieval' "
            "FROM agent_config_snapshots WHERE id='%s';\"" % C.SNAP_ID
        )
        print("验证 retrieval:", out.strip())
        print("\n[WARN] 改完不要点「构建/发布」")
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env bash
set -euo pipefail

docker exec docker-db_postgres-1 psql -U postgres -d dify -c "
SELECT config_snapshot::json -> 'knowledge' -> 'sets' -> 0 -> 'retrieval'
FROM agent_config_snapshots
WHERE id = '01a0135c-320e-79dc-97a0-94da9b904c13';
"

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 8：开启 Rerank 后的命中测试（修改后）。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


QUERY = "法定假日加班的制度是什么，加班费怎么算？"


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        r = d.post("/datasets/%s/hit-testing" % C.DS_ID, body={"query": QUERY})
        pp("hit-testing after rerank", r)
        records = r.get("json")
        if not isinstance(records, list):
            records = (records or {}).get("records") or []
        print("\n--- 前 5 条（分数为 Rerank relevance_score）---")
        for i, rec in enumerate(records[:5], 1):
            seg = (rec.get("segment") or {}).get("content") or ""
            print(i, rec.get("score"), seg[:120].replace("\n", " "))
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

法定假日加班的制度是什么，加班费怎么算？

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 9：新开 Agent 对话验收（修改后）。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify


QUERY = "法定假日加班的制度是什么，加班费怎么算？"


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        payload = {
            "inputs": {},
            "query": QUERY,
            "response_mode": "streaming",
            "files": [],
            "conversation_id": "",
        }
        r = d.post(
            "/agent/%s/chat-messages" % C.AGENT_ID,
            body=payload,
            extra="-m 420 --max-time 420",
            t=450,
        )
        print("HTTP", r.get("code"))
        text = r.get("text") or ""
        print(text[-1500:] if len(text) > 1500 else text)
        if r.get("stderr"):
            print("[stderr]", r["stderr"][:400])
        return 0 if r.get("code") == "200" else 1
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

Task tasks.ops_trace_task.process_trace_tasks ... Processing trace tasks success

#!/usr/bin/env bash
set -euo pipefail

curl -s -u 'pk-lf-84a76d89-b6b4-4943-b4ea-2dee48d4cda7:sk-lf-24aa6d73-ed36-4447-9eb6-3ebd446fb298' \
  'http://127.0.0.1:3000/api/public/traces?limit=8'

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 1：确认现状（Rerank 开启前的基线）。"""
from __future__ import annotations

import json
import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


def main() -> int:
    print("=== 1/4 Xinference 容器与内存（在 207 上 ssh 执行 docker 命令）===")
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        for cmd in [
            "docker ps --filter name=%s --format '{{.Names}} {{.Status}} {{.Ports}}'" % C.XF_CONTAINER,
            "docker stats %s --no-stream --format '{{.MemUsage}} CPU={{.CPUPerc}}'" % C.XF_CONTAINER,
            "free -h",
        ]:
            out, err = d.sh(cmd)
            print(out or err)

        print("\n=== 2/4 Xinference 已加载模型（应只有 bge-m3 embedding）===")
        out, err = d.sh(
            "curl -s -H 'Authorization: Bearer %s' %s/v1/models" % (C.XF_KEY, C.XF_URL)
        )
        print(out or err)
        try:
            models = json.loads(out)
            for m in models.get("data", models if isinstance(models, list) else []):
                if isinstance(m, dict):
                    print(" -", m.get("id"), m.get("model_type"), m.get("accelerators"))
        except Exception:
            pass

        print("\n=== 3/4 登录 Dify，读知识库 retrieval_model ===")
        ok, _, csrf = d.login()
        print("login ok:", ok, "csrf:", bool(csrf))
        r = d.get("/datasets/%s" % C.DS_ID)
        pp("dataset retrieval_model_dict", r)
        retr = (r.get("json") or {}).get("retrieval_model_dict") or {}
        if retr.get("reranking_enable"):
            print("[WARN] 知识库已开 Rerank，不是「修改前」基线")

        print("\n=== 4/4 Agent 快照 knowledge.sets[0].retrieval + rerank 模型列表 ===")
        sql = (
            "docker exec docker-db_postgres-1 psql -U postgres -d dify -t -A -c "
            "\"SELECT config_snapshot::jsonb -> 'knowledge' -> 'sets' -> 0 -> 'retrieval' "
            "FROM agent_config_snapshots WHERE id='%s';\"" % C.SNAP_ID
        )
        out, err = d.sh(sql)
        print("agent retrieval:", out.strip() or err)
        r2 = d.get("/workspaces/current/models/model-types/rerank")
        pp("model-types/rerank（期望 data 为空）", r2)
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 2：开启 Rerank 前的命中测试（对照基线）。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


QUERY = "法定假日加班的制度是什么"


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        r = d.post(
            "/datasets/%s/hit-testing" % C.DS_ID,
            body={
                "query": QUERY,
                "retrieval_model": {
                    "search_method": "semantic_search",
                    "reranking_enable": False,
                    "top_k": 5,
                    "score_threshold_enabled": False,
                },
            },
        )
        pp("hit-testing before rerank", r)
        records = r.get("json")
        if not isinstance(records, list):
            records = (records or {}).get("records") or []
        print("\n--- 前 5 条 ---")
        for i, rec in enumerate(records[:5], 1):
            seg = (rec.get("segment") or {}).get("content") or ""
            print(i, rec.get("score"), seg[:80].replace("\n", " "))
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 3.2 路径 B：把 Dify 的 Xinference URL 改到 GPU 机（修改后）。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


def patch_credentials(d: Dify, model: str, model_type: str, server_url: str, api_key: str) -> None:
    body = {
        "model": model,
        "model_type": model_type,
        "credentials": {
            "server_url": server_url,
            "model_uid": model,
            "api_key": api_key,
            "invoke_timeout": "60",
            "max_retries": "3",
        },
        "name": model,
    }
    path = "/workspaces/current/model-providers/%s/models/credentials" % C.XINFERENCE_PROVIDER
    r = d.post(path, body)
    pp("POST %s %s" % (model_type, model), r)
    if r.get("code") not in ("200", "201"):
        r = d.req("PUT", path, body=body)
        pp("PUT %s %s" % (model_type, model), r)


def main() -> int:
    gpu_key = C.XF_KEY  # 新起 GPU 容器时换成 116 上生成的 xf- Key
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        print("=== 修改前 URL（207 CPU）===")
        print("embedding/rerank server_url:", C.XF_URL)
        print("\n=== 修改后 URL（116 GPU）===")
        print("embedding/rerank server_url:", C.XF_GPU_URL)
        patch_credentials(d, C.EMBED_MODEL, "text-embedding", C.XF_GPU_URL, gpu_key)
        patch_credentials(d, C.RERANK_MODEL, "rerank", C.XF_GPU_URL, gpu_key)
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 5：把 Rerank 模型加进 Dify 供应商凭据。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        body = {
            "model": C.RERANK_MODEL,
            "model_type": "rerank",
            "credentials": {
                "server_url": C.XF_URL,
                "model_uid": C.RERANK_MODEL,
                "api_key": C.XF_KEY,
                "invoke_timeout": "60",
                "max_retries": "3",
            },
            "name": C.RERANK_MODEL,
        }
        r = d.post(
            "/workspaces/current/model-providers/%s/models/credentials" % C.XINFERENCE_PROVIDER,
            body,
        )
        pp("add rerank cred", r)
        r2 = d.get("/workspaces/current/models/model-types/rerank")
        pp("rerank models", r2)
        return 0 if r.get("code") in ("200", "201") else 1
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 6 · 修改前：只读知识库 retrieval_model（不要 PATCH）。"""
from __future__ import annotations

import json
import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        r = d.get("/datasets/%s" % C.DS_ID)
        pp("dataset before patch", r)
        retr = (r.get("json") or {}).get("retrieval_model_dict") or {}
        print("\n期望修改前关键字段：")
        print(json.dumps({
            "reranking_enable": retr.get("reranking_enable"),
            "top_k": retr.get("top_k"),
            "reranking_mode": retr.get("reranking_mode"),
            "reranking_model": retr.get("reranking_model"),
        }, ensure_ascii=False, indent=2))
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 6 · 修改后：PATCH 知识库，开启 Rerank，Top-K=5。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


def build_retrieval_after() -> dict:
    return {
        "search_method": "semantic_search",
        "reranking_enable": True,
        "reranking_mode": "reranking_model",
        "reranking_model": {
            "reranking_provider_name": C.XINFERENCE_PROVIDER,
            "reranking_model_name": C.RERANK_MODEL,
        },
        "top_k": 5,
        "score_threshold_enabled": False,
        "score_threshold": None,
    }


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        ds_retrieval = build_retrieval_after()
        r = d.req("PATCH", "/datasets/%s" % C.DS_ID, body={"retrieval_model": ds_retrieval})
        print("HTTP", r.get("code"))
        j = r.get("json") or {}
        print("retrieval_model_dict:", j.get("retrieval_model_dict"))
        pp("patch dataset", r)
        return 0 if r.get("code") == "200" else 1
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 7 · 修改前：备份 Agent 快照/草稿 retrieval（只读）。"""
from __future__ import annotations

import json
import sys
from pathlib import Path

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify


def load_cfg(d: Dify, table: str, cid: str) -> dict:
    out, err = d.sh(
        "docker exec docker-db_postgres-1 psql -U postgres -d dify -t -A -c "
        "\"SELECT config_snapshot FROM %s WHERE id='%s'\"" % (table, cid)
    )
    raw = out.strip()
    if not raw:
        raise RuntimeError(err or "empty config_snapshot")
    return json.loads(raw)


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        snap = load_cfg(d, "agent_config_snapshots", C.SNAP_ID)
        draft = load_cfg(d, "agent_config_drafts", C.DRAFT_ID)
        bak_dir = Path(__file__).resolve().parent / "backup"
        bak_dir.mkdir(exist_ok=True)
        for name, cfg in [("snap_before", snap), ("draft_before", draft)]:
            p = bak_dir / ("专题十三_07_%s.json" % name)
            p.write_text(json.dumps(cfg, ensure_ascii=False, indent=2), encoding="utf-8")
            retr = cfg.get("knowledge", {}).get("sets", [{}])[0].get("retrieval", {})
            print(name, "retrieval:", json.dumps(retr, ensure_ascii=False))
            print("  -> 已备份", p)
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 7 · 修改后：同步改 Agent 快照与草稿的 knowledge.sets[0].retrieval。"""
from __future__ import annotations

import json
import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify


def load_cfg(d: Dify, table: str, cid: str) -> dict:
    out, err = d.sh(
        "docker exec docker-db_postgres-1 psql -U postgres -d dify -t -A -c "
        "\"SELECT config_snapshot FROM %s WHERE id='%s'\"" % (table, cid)
    )
    raw = out.strip()
    if not raw:
        raise RuntimeError(err or "empty config_snapshot")
    return json.loads(raw)


def save_cfg(d: Dify, table: str, cid: str, cfg: dict, tag: str) -> None:
    sql = "UPDATE %s SET config_snapshot = $cfg$%s$cfg$ WHERE id='%s';" % (
        table,
        json.dumps(cfg, ensure_ascii=False),
        cid,
    )
    remote = "/tmp/upd_%s.sql" % tag
    with d.sftp.open(remote, "w") as f:
        f.write(sql)
    out, err = d.sh("docker exec -i docker-db_postgres-1 psql -U postgres -d dify < %s" % remote)
    print(table, (out or err).strip())


def enable_agent_rerank(cfg: dict) -> dict:
    retr = cfg["knowledge"]["sets"][0].setdefault("retrieval", {})
    retr["top_k"] = 5
    retr["reranking_enable"] = True
    retr["reranking_mode"] = "reranking_model"
    retr["reranking_model"] = {
        "provider": C.XINFERENCE_PROVIDER,
        "model": C.RERANK_MODEL,
    }
    return cfg


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        snap = load_cfg(d, "agent_config_snapshots", C.SNAP_ID)
        draft = load_cfg(d, "agent_config_drafts", C.DRAFT_ID)
        save_cfg(d, "agent_config_snapshots", C.SNAP_ID, enable_agent_rerank(snap), "snap_rerank")
        save_cfg(d, "agent_config_drafts", C.DRAFT_ID, enable_agent_rerank(draft), "draft_rerank")
        out, _ = d.sh(
            "docker exec docker-db_postgres-1 psql -U postgres -d dify -t -A -c "
            "\"SELECT config_snapshot::json -> 'knowledge' -> 'sets' -> 0 -> 'retrieval' "
            "FROM agent_config_snapshots WHERE id='%s';\"" % C.SNAP_ID
        )
        print("验证 retrieval:", out.strip())
        print("\n[WARN] 改完不要点「构建/发布」")
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 8：开启 Rerank 后的命中测试（修改后）。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify, pp


QUERY = "法定假日加班的制度是什么，加班费怎么算？"


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        r = d.post("/datasets/%s/hit-testing" % C.DS_ID, body={"query": QUERY})
        pp("hit-testing after rerank", r)
        records = r.get("json")
        if not isinstance(records, list):
            records = (records or {}).get("records") or []
        print("\n--- 前 5 条（分数为 Rerank relevance_score）---")
        for i, rec in enumerate(records[:5], 1):
            seg = (rec.get("segment") or {}).get("content") or ""
            print(i, rec.get("score"), seg[:120].replace("\n", " "))
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题十三 · 步骤 9：新开 Agent 对话验收（修改后）。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify


QUERY = "法定假日加班的制度是什么，加班费怎么算？"


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        payload = {
            "inputs": {},
            "query": QUERY,
            "response_mode": "streaming",
            "files": [],
            "conversation_id": "",
        }
        r = d.post(
            "/agent/%s/chat-messages" % C.AGENT_ID,
            body=payload,
            extra="-m 420 --max-time 420",
            t=450,
        )
        print("HTTP", r.get("code"))
        text = r.get("text") or ""
        print(text[-1500:] if len(text) > 1500 else text)
        if r.get("stderr"):
            print("[stderr]", r["stderr"][:400])
        return 0 if r.get("code") == "200" else 1
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env bash
set -euo pipefail

-H 'Authorization: Bearer xf-K8H3LiS27H4188XdOyWJj74MWMOlIR9tZbiybDbxbSas7eAI'

RuntimeError: Failed to launch model, detail: API keys can only access model query and inference endpoints

#!/usr/bin/env bash
set -euo pipefail

xinference login -e http://127.0.0.1:9997 --username admin --password admin1234
xinference launch ...

RuntimeError: Cannot find access token, please login first!

ValueError: The parameter `n_gpu` must be greater than 0 and not greater than the number of GPUs: 0 on the machine.

#!/usr/bin/env bash
set -euo pipefail

# 错误：无 GPU 机器上传 0 会被拒绝
xinference launch ... --n-gpu 0

# 正确：207 CPU，省略该参数
xinference launch -e http://127.0.0.1:9997 -n bge-reranker-v2-m3 -t rerank -u bge-reranker-v2-m3

#!/usr/bin/env bash
set -euo pipefail

xinference launch -e http://127.0.0.1:9997 \
  -n bge-reranker-v2-m3 -t rerank -u bge-reranker-v2-m3 \
  --n-gpu auto
# 或 --n-gpu 1

#!/usr/bin/env bash
set -euo pipefail

xinference login ...
xinference registrations -e http://127.0.0.1:9997 -t rerank

"reranking_provider_name": request.retrieval.reranking_model.provider,
"reranking_model_name": request.retrieval.reranking_model.model,