feat: 完成 issue #44 ④ 本地 70B 模型接入与推理封装

2026-08-05 01:56:59 +08:00
parent f5d2294ee4
commit 169492a6db
4 changed files with 273 additions and 0 deletions
@@ -0,0 +1,114 @@
+# -*- coding: utf-8 -*-
+"""推理后端实现 —— 本地 70B 模型接入与推理封装（issue #44）。
+
+在 `gateway.InferenceBackend` 抽象之上交付**真实可用的本地后端**：
+- OpenAI 兼容接口（vLLM / TGI 等本地推理服务，`/v1/chat/completions`），
+  仅用标准库 urllib，无第三方依赖；
+- 参数化：endpoint / model / timeout / max_tokens / temperature / context 引用注入；
+- **数据不出厂**（PRD 5.4）：敏感/核心内容走本地后端，云端仅接收脱敏内容；
+- 未配置 endpoint 时进入 dry-run 占位模式（保持与旧 LocalBackend 一致的
+  可测试行为，供端到端演示与联调）。
+
+业务代码只依赖 `gateway.InferenceBackend.generate(prompt, context)`，
+切换后端 = 换实现（见 `LLMGateway(local=...)`）。
+"""
+from __future__ import annotations
+
+import json
+import time
+import urllib.request
+from typing import Optional, Sequence
+
+from .gateway import InferenceBackend
+
+
+class Local70BBackend(InferenceBackend):
+    """本地 70B 推理后端（OpenAI 兼容 vLLM/TGI，参数化）。"""
+
+    name = "local-70b"
+
+    def __init__(
+        self,
+        endpoint: str = "",
+        model: str = "iaop-local-70b",
+        timeout_seconds: float = 60.0,
+        max_tokens: int = 1024,
+        temperature: float = 0.1,
+        echo_context: bool = True,
+    ) -> None:
+        self.endpoint = (endpoint or "").rstrip("/")
+        self.model = model
+        self.timeout = float(timeout_seconds)
+        self.max_tokens = int(max_tokens)
+        self.temperature = float(temperature)
+        self.echo_context = echo_context
+
+    # ------------------------------------------------------------------
+    def generate(self, prompt: str, context: Sequence[str]) -> str:
+        """根据 prompt 与 RAG 上下文生成回答。
+
+        - 未配置 endpoint：dry-run 占位（回显 prompt 前 40 字符 + 来源引用）；
+        - 已配置：调用本地 OpenAI 兼容服务（/v1/chat/completions）。
+        """
+        if not self.endpoint:
+            return self._dry_run(prompt, context)
+
+        payload = {
+            "model": self.model,
+            "messages": [
+                {"role": "system", "content": self._system_prompt(context)},
+                {"role": "user", "content": prompt},
+            ],
+            "max_tokens": self.max_tokens,
+            "temperature": self.temperature,
+        }
+        body = self._post_json("/v1/chat/completions", payload)
+        try:
+            return body["choices"][0]["message"]["content"]
+        except (KeyError, IndexError, TypeError):
+            raise RuntimeError(
+                f"本地推理服务响应格式异常: {str(body)[:200]}")
+
+    # ------------------------------------------------------------------
+    def _system_prompt(self, context: Sequence[str]) -> str:
+        """把 RAG 引用注入 system 提示（引用溯源，PRD 5.4）。"""
+        refs = "\n".join(f"- {c}" for c in (context or []))
+        base = "你是工业 AI 助手。回答须基于给定资料并标注来源。"
+        return f"{base}\n参考资料：\n{refs}" if refs else base
+
+    def _dry_run(self, prompt: str, context: Sequence[str]) -> str:
+        head = f"[本地70B占位] {prompt[:40]}"
+        if self.echo_context:
+            for i, src in enumerate(context[:3], 1):
+                head += f"\n[来源: {src}]"
+        return head
+
+    def _post_json(self, path: str, payload: dict) -> dict:
+        """向后端推理服务发起 JSON POST（标准库 urllib）。"""
+        url = self.endpoint + path
+        data = json.dumps(payload).encode("utf-8")
+        req = urllib.request.Request(
+            url, data=data,
+            headers={"Content-Type": "application/json"})
+        with urllib.request.urlopen(req, timeout=self.timeout) as resp:
+            raw = resp.read().decode("utf-8")
+        return json.loads(raw) if raw else {}
+
+    def health(self) -> dict:
+        """后端健康信息（本地推理服务可探测 /health）。"""
+        base = {
+            "backend": self.name, "model": self.model,
+            "endpoint": self.endpoint or "(dry-run)",
+        }
+        if not self.endpoint:
+            base["status"] = "dry-run"
+            return base
+        try:
+            started = time.monotonic()
+            with urllib.request.urlopen(
+                    self.endpoint + "/health", timeout=self.timeout) as resp:
+                base["status"] = "ok" if resp.status == 200 else f"http-{resp.status}"
+            base["latency_ms"] = round((time.monotonic() - started) * 1000, 2)
+        except Exception as exc:  # noqa: BLE001 - 健康探测失败仅记录
+            base["status"] = f"error: {exc}"
+        return base