merge: bge-reranker service from GPU server

.gitignore

@@ -0,0 +1,7 @@
+ah.pid
+nohup.out
+__pycache__/
+*.pyc
+workers/__pycache__/
+logs/
+files/

README.md

@@ -1,2 +1,79 @@
-# bge-reranker
+# BGE-Reranker Service
 
+BGE-Reranker-V2-M3 文本重排序服务，用于 RAG 检索后的精排阶段。
+
+## 特性
+
+- 模型: BAAI/bge-reranker-v2-m3 (2.2GB)
+- 推理: FP16, GPU 加速
+- 延迟: ~45ms (模型加载后)
+- 最大文档数: 100条/请求
+
+## 部署
+
+```bash
+cd /data/ymq/bge-reranker
+bash build.sh deploy   # 启动
+bash build.sh stop     # 停止
+bash build.sh status   # 状态
+```
+
+## API
+
+### GET /api/status
+服务状态。
+
+### POST /api/rerank
+```json
+{
+    "query": "安装步骤",
+    "documents": ["打开包装箱", "拧下螺丝", "注意事项"],
+    "top_k": 3
+}
+```
+
+返回:
+```json
+{
+    "status": "SUCCEEDED",
+    "query": "安装步骤",
+    "ranked_docs": [
+        {"doc": "拧下螺丝", "score": 0.164},
+        {"doc": "打开包装箱", "score": 0.065},
+        {"doc": "注意事项", "score": 0.017}
+    ],
+    "total": 3,
+    "returned": 3,
+    "elapsed": 0.0456
+}
+```
+
+## 在 RAG 中的角色
+
+```
+查询 → CLIP embedding → VDB 向量召回 top-20
+                      → FalkorDB 图扩展
+                      → 合并
+                      → BGE-Rerank (本服务) → top-5
+                      → LLM 生成答案
+```
+
+## 离线模型下载
+
+```bash
+# 使用 huggingface-cli
+huggingface-cli download BAAI/bge-reranker-v2-m3 --local-dir /data/ymq/models/BAAI/bge-reranker-v2-m3
+
+# 或使用 hf
+hf download BAAI/bge-reranker-v2-m3 --local-dir /data/ymq/models/BAAI/bge-reranker-v2-m3
+```
+
+## 配置
+
+- 端口: 9090
+- GPU: 2
+- 虚拟环境: /data/ymq/wan22-service/py3
+
+## Git
+
+git@git.opencomputing.cn:yumoqing/bge-reranker.git

ah.py

@@ -0,0 +1,6 @@
+# -*- coding:utf-8 -*-
+from ahserver.webapp import webapp
+from init import load_bge_reranker
+
+if __name__ == '__main__':
+    webapp(load_bge_reranker)

build.sh

@@ -0,0 +1,57 @@
+#!/usr/bin/env bash
+# BGE-Reranker Service
+set -e
+cd "$(dirname "$0")"
+
+SERVICE_NAME="bge-reranker"
+PORT=9090
+GPU=2
+PY=/data/ymq/wan22-service/py3/bin/python
+action="${1:-status}"
+
+case "$action" in
+    deploy|update)
+        echo "=== $SERVICE_NAME Deploy (GPU $GPU, port $PORT) ==="
+        if [ -f ah.pid ] && kill -0 $(cat ah.pid) 2>/dev/null; then
+            kill $(cat ah.pid) 2>/dev/null || true; sleep 2
+        fi
+        if [ -d .git ] && [ -f .git/HEAD ]; then
+            git pull origin master 2>/dev/null || true
+        fi
+        mkdir -p logs files wwwroot
+        export PYTHONPATH="$(pwd)"
+        export CUDA_VISIBLE_DEVICES=$GPU
+        nohup $PY ah.py > nohup.out 2>&1 &
+        echo $! > ah.pid
+        echo "Started PID $(cat ah.pid) on port $PORT (GPU $GPU)"
+        sleep 5
+        if curl -s http://localhost:$PORT/api/status > /dev/null 2>&1; then
+            echo "Service healthy"
+        else
+            echo "WARNING: not responding, check nohup.out"
+            tail -20 nohup.out
+        fi
+        ;;
+    stop)
+        if [ -f ah.pid ]; then
+            kill $(cat ah.pid) 2>/dev/null || true; rm -f ah.pid; echo "Stopped"
+        else echo "Not running"; fi
+        ;;
+    start)
+        mkdir -p logs files wwwroot
+        export PYTHONPATH="$(pwd)"; export CUDA_VISIBLE_DEVICES=$GPU
+        nohup $PY ah.py > nohup.out 2>&1 &
+        echo $! > ah.pid; echo "Started PID $(cat ah.pid)"
+        ;;
+    status)
+        echo "=== $SERVICE_NAME Status ==="
+        if [ -f ah.pid ] && kill -0 $(cat ah.pid) 2>/dev/null; then
+            echo "Process: running (PID $(cat ah.pid))"
+        else echo "Process: not running"; fi
+        echo "Port: $PORT, GPU: $GPU"
+        if curl -s --max-time 3 http://localhost:$PORT/api/status > /dev/null 2>&1; then
+            echo "HTTP: OK"
+        else echo "HTTP: not responding"; fi
+        ;;
+    *) echo "Usage: $0 {deploy|update|stop|start|status}"; exit 1 ;;
+esac

conf/config.json

@@ -0,0 +1,25 @@
+{
+    "password_key": "BgeReranker2026Key",
+    "filesroot": "$[workdir]$/files",
+    "logger": {
+        "name": "bge-reranker",
+        "levelname": "info",
+        "logfile": "$[workdir]$/logs/bge-reranker.log"
+    },
+    "website": {
+        "paths": [["$[workdir]$/wwwroot", ""]],
+        "client_max_size": 10485760,
+        "host": "0.0.0.0",
+        "port": 9090,
+        "coding": "utf-8",
+        "indexes": ["index.html"],
+        "startswiths": [
+            {"leading": "/api/status", "registerfunction": "status"},
+            {"leading": "/api/rerank", "registerfunction": "rerank"}
+        ],
+        "processors": [
+            [".tmpl", "tmpl"], [".app", "app"], [".ui", "bui"],
+            [".dspy", "dspy"], [".md", "md"]
+        ]
+    }
+}

init.py

@@ -0,0 +1,55 @@
+# -*- coding:utf-8 -*-
+from traceback import format_exc
+from ahserver.serverenv import ServerEnv
+from appPublic.registerfunction import RegisterFunction
+from appPublic.log import debug, exception
+import json
+
+async def status_handler(request, params_kw, *args, **kwargs):
+    """Status endpoint"""
+    return json.dumps({
+        "service": "bge-reranker",
+        "model": "BAAI/bge-reranker-v2-m3",
+        "model_path": "/data/ymq/models/BAAI/bge-reranker-v2-m3",
+        "endpoints": ["/api/status", "/api/rerank"]
+    }, indent=2, ensure_ascii=False)
+
+
+async def rerank_handler(request, params_kw, *args, **kwargs):
+    """Rerank endpoint"""
+    try:
+        query = params_kw.get("query")
+        documents = params_kw.get("documents", [])
+        top_k = params_kw.get("top_k", 5)
+        
+        if not query or not documents:
+            return json.dumps({"error": "query and documents required"})
+        
+        if not isinstance(documents, list):
+            return json.dumps({"error": "documents must be a list"})
+        
+        from workers.bge_reranker import rerank
+        import time
+        start = time.time()
+        result = rerank(query, documents, top_k)
+        elapsed = round(time.time() - start, 4)
+        
+        return json.dumps({
+            "status": "SUCCEEDED",
+            "query": query,
+            "ranked_docs": result["ranked_docs"],
+            "total": result["total"],
+            "returned": result["returned"],
+            "elapsed": elapsed
+        }, ensure_ascii=False)
+    except Exception as e:
+        exception(f"{e}, {format_exc()}")
+        return json.dumps({"error": str(e)})
+
+
+def load_bge_reranker():
+    """Register API handlers"""
+    env = ServerEnv()
+    rf = RegisterFunction()
+    rf.register("status", status_handler)
+    rf.register("rerank", rerank_handler)

workers/bge_reranker.py

@@ -0,0 +1,60 @@
+# -*- coding:utf-8 -*-
+"""BGE-Reranker-M3 lazy-loading wrapper."""
+import os
+import json
+import torch
+from FlagEmbedding import FlagReranker
+
+_model = None
+_lock = False
+
+MODEL_PATH = "/data/ymq/models/BAAI/bge-reranker-v2-m3"
+DEVICE = "cuda:0"
+
+def get_model():
+    global _model, _lock
+    if _model is None and not _lock:
+        _lock = True
+        try:
+            _model = FlagReranker(MODEL_PATH, use_fp16=True, device=DEVICE)
+        except Exception as e:
+            _lock = False
+            raise RuntimeError(f"Failed to load BGE-Reranker: {e}")
+    return _model
+
+
+def rerank(query: str, documents: list, top_k: int = 5) -> dict:
+    """Rerank documents by relevance to query."""
+    model = get_model()
+    
+    # Create pairs: [[query, doc1], [query, doc2], ...]
+    pairs = [[query, doc] for doc in documents]
+    
+    # Compute scores
+    scores = model.compute_score(pairs, normalize=True)
+    
+    if isinstance(scores, float):
+        scores = [scores]
+    
+    # Pair with documents and sort
+    ranked = list(zip(scores, documents))
+    ranked.sort(key=lambda x: -x[0])
+    
+    # Return top_k
+    results = [
+        {"doc": doc, "score": round(score, 6)}
+        for score, doc in ranked[:top_k]
+    ]
+    
+    return {"ranked_docs": results, "total": len(documents), "returned": len(results)}
+
+
+def health_check():
+    """Check model status."""
+    model = get_model()
+    return {
+        "model": "BAAI/bge-reranker-v2-m3",
+        "device": DEVICE,
+        "loaded": model is not None,
+        "model_path": MODEL_PATH
+    }