{"id":174,"date":"2026-05-20T07:02:37","date_gmt":"2026-05-19T23:02:37","guid":{"rendered":"https:\/\/zoran.090702.xyz\/?p=174"},"modified":"2026-05-20T07:02:37","modified_gmt":"2026-05-19T23:02:37","slug":"%e6%8a%8a8b%e5%b0%8f%e6%a8%a1%e5%9e%8b%e4%bb%8e53%e6%8b%89%e5%88%b099%ef%bc%9aforge-%e5%bc%80%e6%ba%90%e9%a1%b9%e7%9b%ae%e5%ae%9e%e6%88%98%e8%a7%a3%e6%9e%90","status":"publish","type":"post","link":"https:\/\/zoran.090702.xyz\/?p=174","title":{"rendered":"\u628a8B\u5c0f\u6a21\u578b\u4ece53%\u62c9\u523099%\uff1aForge \u5f00\u6e90\u9879\u76ee\u5b9e\u6218\u89e3\u6790"},"content":{"rendered":"<p>\u524d\u4e24\u5929\u901b Hacker News\uff0c\u770b\u5230\u4e00\u4e2a\u9879\u76ee\u8ba9\u6211\u773c\u524d\u4e00\u4eae\u2014\u2014**Forge**\uff0c\u4e00\u4e2a\u7ed9\u672c\u5730\u5c0f\u6a21\u578b\u52a0\u62a4\u680f\uff08Guardrails\uff09\u7684\u5f00\u6e90\u6846\u67b6\u3002<\/p>\n<p>\u5b83\u7684\u53e3\u53f7\u5f88\u72c2\u5984\uff1a\u628a 8B \u53c2\u6570\u7684\u5c0f\u6a21\u578b\u5728 agent \u4efb\u52a1\u4e0a\u7684\u51c6\u786e\u7387\u4ece 53% \u5e72\u5230 99%\u3002<\/p>\n<p>\u6211\u4e00\u5f00\u59cb\u4e0d\u4fe1\u30028B \u6a21\u578b\u4ec0\u4e48\u6c34\u5e73\u5927\u5bb6\u5fc3\u91cc\u6709\u6570\uff0c\u8dd1\u8dd1\u804a\u5929\u8fd8\u884c\uff0c\u771f\u8981\u5b83\u8c03\u5de5\u5177\u3001\u8d70\u591a\u6b65\u6d41\u7a0b\uff0c\u7ecf\u5e38\u534a\u8def\u8dd1\u504f\u3002\u7ed3\u679c\u770b\u4e86\u4f5c\u8005\u7684\u57fa\u51c6\u6d4b\u8bd5\u2014\u2014Ministral-3 8B Instruct Q8 \u8dd1 llama-server\uff0c\u5728 26 \u4e2a\u573a\u666f\u7684\u8bc4\u6d4b\u96c6\u4e0a\u5e72\u5230\u4e86 **86.5%**\uff0c\u6700\u96be\u7684\u90a3\u4e00\u6863\u4e5f\u6709 76%\u3002\u914d\u5408\u63a8\u7406\u65f6\u6269\u5c55\uff08test-time compute scaling\uff09\uff0c\u76f4\u63a5\u51b2\u5230 **99%**\u3002<\/p>\n<p>\u8fd9\u6570\u5b57\u592a\u79bb\u8c31\u4e86\uff0c\u5fc5\u987b\u62c6\u5f00\u6765\u770b\u770b\u3002<\/p>\n<p>## \u5230\u5e95\u600e\u4e48\u505a\u5230\u7684\uff1f<\/p>\n<p>Forge \u7684\u6838\u5fc3\u601d\u8def\u5176\u5b9e\u4e0d\u590d\u6742\uff1a**\u4e0d\u8981\u8ba9\u6a21\u578b\u81ea\u5df1\u51b3\u5b9a\u600e\u4e48\u5e72\u6d3b\uff0c\u7ed9\u5b83\u5957\u4e0a\u89c4\u77e9\u3002**<\/p>\n<p>\u5177\u4f53\u6765\u8bf4\u6709\u4e09\u5c42\uff1a<\/p>\n<p>**1. Rescue Parsing\uff08\u6551\u573a\u5f0f\u89e3\u6790\uff09**<\/p>\n<p>\u6a21\u578b\u8f93\u51fa\u7684\u6700\u5927\u95ee\u9898\u662f\u683c\u5f0f\u8dd1\u504f\u3002\u4f60\u8ba9\u5b83\u8f93\u51fa JSON\uff0c\u5b83\u7ed9\u4f60\u6765\u4e00\u6bb5\u6563\u6587\u3002Forge \u7684\u505a\u6cd5\u662f\u2014\u2014\u4e0d\u7ba1\u6a21\u578b\u8f93\u51fa\u4ec0\u4e48\uff0c\u90fd\u7528\u53e6\u4e00\u4e2a\u89e3\u6790\u5668\u53bb\u635e\u6709\u7528\u7684\u4e1c\u897f\u3002\u8f93\u51fa\u4e0d\u5408\u6cd5\uff1f\u4e0d\u62a5\u9519\uff0c\u76f4\u63a5\u91cd\u8bd5\u52a0\u63d0\u793a\u4fee\u6b63\u3002<\/p>\n<p>**2. Step Enforcement\uff08\u6b65\u9aa4\u5f3a\u5236\u6267\u884c\uff09**<\/p>\n<p>\u591a\u6b65 agent \u6d41\u7a0b\u4e2d\u6700\u5e38\u89c1\u7684\u7ffb\u8f66\u70b9\u662f\uff1a\u6a21\u578b\u8df3\u6b65\u9aa4\u3002Forge \u5728\u7cfb\u7edf prompt \u91cc\u5d4c\u4e86\u6b65\u9aa4\u7ea6\u675f\uff0c\u6bcf\u4e00\u6b65\u53ea\u5141\u8bb8\u6a21\u578b\u505a\u5f53\u524d\u8be5\u505a\u7684\u4e8b\uff0c\u505a\u5b8c\u518d\u653e\u884c\u5230\u4e0b\u4e00\u6b65\u3002\u5f88\u50cf\u8d70\u94a2\u7d22\u4e0b\u9762\u7684\u5b89\u5168\u7f51\u2014\u2014\u5b83\u80fd\u8d70\u6b6a\uff0c\u4f46\u6b6a\u4e86\u4f1a\u88ab\u6276\u56de\u6765\u3002<\/p>\n<p>**3. Context Compaction\uff08\u4e0a\u4e0b\u6587\u538b\u7f29\uff09**<\/p>\n<p>\u672c\u5730\u6a21\u578b\u6700\u5934\u75bc\u7684\u662f\u4e0a\u4e0b\u6587\u7a97\u53e3\u3002\u51e0\u8f6e\u5de5\u5177\u8c03\u7528\u4e0b\u6765\uff0c\u5386\u53f2\u5c31\u6491\u7206\u4e86\u3002Forge \u505a\u4e86\u5206\u5c42\u538b\u7f29\uff1a\u65e7\u5bf9\u8bdd\u5148\u6458\u8981\uff0c\u518d\u6309\u91cd\u8981\u6027\u6392\u5e8f\u4e22\u5f03\uff0c\u786e\u4fdd\u7a97\u53e3\u59cb\u7ec8\u7559\u7ed9\u6700\u91cd\u8981\u7684\u5185\u5bb9\u3002<\/p>\n<p>## \u4e0a\u624b\u4f53\u9a8c<\/p>\n<p>Forge \u63d0\u4f9b\u4e86\u4e09\u79cd\u63a5\u5165\u65b9\u5f0f\uff0c\u6309\u9700\u6c42\u9009\u5c31\u884c\uff1a<\/p>\n<p>### \u65b9\u5f0f\u4e00\uff1aWorkflowRunner\uff08\u81ea\u5e26\u6d41\u7a0b\u5f15\u64ce\uff09<\/p>\n<p>\u9002\u5408\u4ece\u5934\u5f00\u59cb\u642d agent \u7684\u3002\u5b9a\u4e49\u597d\u5de5\u5177\u5217\u8868\uff0c\u9009\u4e2a\u540e\u7aef\uff0cForge \u5e2e\u4f60\u7ba1\u5168\u5957\u2014\u2014prompt \u7ec4\u88c5\u3001\u5de5\u5177\u6267\u884c\u3001\u4e0a\u4e0b\u6587\u7ba1\u7406\u3001\u62a4\u680f\u5168\u90e8\u5185\u7f6e\u3002<\/p>\n<p>&#8220;`python<br \/>\nfrom forge import WorkflowRunner<br \/>\nfrom forge.backends import LlamaCppBackend<\/p>\n<p># \u8fde\u63a5\u672c\u5730 llama-server<br \/>\nbackend = LlamaCppBackend(base_url=&#8221;http:\/\/localhost:8080&#8243;)<br \/>\nrunner = WorkflowRunner(backend=backend)<\/p>\n<p># \u5b9a\u4e49\u5de5\u5177<br \/>\ndef search_web(query: str) -> str:<br \/>\n    # &#8230; \u4f60\u7684\u641c\u7d22\u903b\u8f91<br \/>\n    return results<\/p>\n<p>def calculate(expr: str) -> float:<br \/>\n    return eval(expr)<\/p>\n<p>runner.add_tool(search_web)<br \/>\nrunner.add_tool(calculate)<\/p>\n<p># \u8dd1\u4e00\u4e2a\u591a\u6b65\u4efb\u52a1<br \/>\nresult = runner.run(&#8220;\u67e5\u8be2\u4eca\u5929\u6bd4\u7279\u5e01\u4ef7\u683c\uff0c\u8ba1\u7b97\u8fc7\u53bb24\u5c0f\u65f6\u6da8\u8dcc\u5e45&#8221;)<br \/>\n&#8220;`<\/p>\n<p>### \u65b9\u5f0f\u4e8c\uff1aGuardrails \u4e2d\u95f4\u4ef6<\/p>\n<p>\u5982\u679c\u4f60\u5df2\u7ecf\u6709\u81ea\u5df1\u7684 agent \u5faa\u73af\u4e86\uff0c\u4e0d\u60f3\u5168\u76d8\u6362\u6389\uff0c\u53ef\u4ee5\u53ea\u6302 Forge \u7684\u62a4\u680f\u5c42\u3002\u5b83\u66b4\u9732\u4e86\u4e00\u5957\u4e2d\u95f4\u4ef6\u63a5\u53e3\uff0c\u63d2\u5230\u4f60\u73b0\u6709\u7684\u7f16\u6392\u903b\u8f91\u91cc\u5c31\u884c\u3002<\/p>\n<p>&#8220;`python<br \/>\nfrom forge.middleware import GuardrailsMiddleware<\/p>\n<p>guard = GuardrailsMiddleware(<br \/>\n    rescue_parsing=True,<br \/>\n    enforce_steps=[&#8220;plan&#8221;, &#8220;execute&#8221;, &#8220;verify&#8221;],<br \/>\n    max_retries=3<br \/>\n)<\/p>\n<p># \u5728\u4f60\u7684\u5faa\u73af\u91cc\u8c03\u7528<br \/>\nfor step in my_pipeline:<br \/>\n    response = model.generate(prompt)<br \/>\n    validated = guard.validate(step, response)<br \/>\n    if not validated.passed:<br \/>\n        response = guard.rescue(response)<br \/>\n&#8220;`<\/p>\n<p>\u5b98\u65b9\u793a\u4f8b\u91cc\u6709\u4e2a\u6587\u4ef6\u53eb `examples\/foreign_loop.py`\uff0c\u5c55\u793a\u7684\u5c31\u662f\u8fd9\u79cd\u7528\u6cd5\u3002<\/p>\n<p>### \u65b9\u5f0f\u4e09\uff1aProxy Server\uff08\u96f6\u4fb5\u5165\uff09<\/p>\n<p>\u8fd9\u662f\u6211\u6700\u559c\u6b22\u7684\u65b9\u5f0f\u3002Forge \u542f\u52a8\u4e00\u4e2a\u517c\u5bb9 OpenAI API \u7684\u4ee3\u7406\u670d\u52a1\uff0c\u4f60\u73b0\u6709\u7684\u5de5\u5177\uff08OpenCode, Continue, aider \u7b49\u7b49\uff09\u5b8c\u5168\u4e0d\u9700\u8981\u6539\u4ee3\u7801\uff0c\u628a base_url \u6307\u5230 Forge \u7684\u4ee3\u7406\u7aef\u53e3\u5c31\u884c\u3002<\/p>\n<p>&#8220;`bash<br \/>\npython -m forge.proxy &#8211;backend http:\/\/localhost:8080 &#8211;port 8081<br \/>\n&#8220;`<\/p>\n<p>\u7136\u540e\u5728\u4f60\u7684\u5ba2\u6237\u7aef\u91cc\uff1a<\/p>\n<p>&#8220;`python<br \/>\nfrom openai import OpenAI<\/p>\n<p>client = OpenAI(<br \/>\n    base_url=&#8221;http:\/\/localhost:8081\/v1&#8243;,<br \/>\n    api_key=&#8221;not-needed&#8221;<br \/>\n)<br \/>\n&#8220;`<\/p>\n<p>\u900f\u660e\u5730\u83b7\u5f97\u62a4\u680f\u52a0\u6301\uff0c\u6a21\u578b\u8868\u73b0\u63d0\u5347\u8089\u773c\u53ef\u89c1\u3002<\/p>\n<p>## \u540e\u7aef\u600e\u4e48\u9009<\/p>\n<p>Forge \u652f\u6301\u56db\u79cd\u540e\u7aef\uff1a<\/p>\n<p>| \u540e\u7aef | \u6027\u80fd | \u4e0a\u624b\u96be\u5ea6 |<br \/>\n|&#8212;&#8212;|&#8212;&#8212;|&#8212;&#8212;&#8212;-|<br \/>\n| llama-server (llama.cpp) | \u2b50\u2b50\u2b50\u2b50\u2b50 | \u2b50\u2b50\u2b50 |<br \/>\n| Ollama | \u2b50\u2b50\u2b50\u2b50 | \u2b50 |<br \/>\n| Llamafile | \u2b50\u2b50\u2b50 | \u2b50\u2b50 |<br \/>\n| Anthropic API | \u2b50\u2b50\u2b50\u2b50\u2b50 (\u4e91) | \u2b50 |<\/p>\n<p>\u57fa\u51c6\u6d4b\u8bd5\u91cc\u6392\u540d\u524d 10 \u7684\u914d\u7f6e\u5168\u8dd1\u5728 **llama-server** \u4e0a\u3002\u5982\u679c\u4f60\u6709 GPU\uff0c\u63a8\u8350\u8fd9\u6761\u8def\u7ebf\uff1a<\/p>\n<p>&#8220;`bash<br \/>\n# \u62c9\u6a21\u578b<br \/>\nwget https:\/\/huggingface.co\/mistralai\/Ministral-3-8B-Instruct-2512-GGUF\/resolve\/main\/Ministral-3-8B-Instruct-2512-Q8_0.gguf<\/p>\n<p># \u542f\u52a8\u670d\u52a1<br \/>\nllama-server -m Ministral-3-8B-Instruct-2512-Q8_0.gguf &#8211;jinja -ngl 999 &#8211;port 8080<br \/>\n&#8220;`<\/p>\n<p>\u6ca1 GPU \u7684\u8bdd Ollama \u6700\u7701\u4e8b\uff1a<\/p>\n<p>&#8220;`bash<br \/>\nollama pull ministral-3:8b-instruct-2512-q4_K_M<br \/>\n&#8220;`<\/p>\n<p>## \u6211\u7684\u5b9e\u9645\u611f\u53d7<\/p>\n<p>\u6211\u62ff\u4e00\u4e2a\u5185\u90e8\u7684\u6570\u636e\u5206\u6790 agent \u8bd5\u4e86\u4e00\u4e0b\u3002\u4e4b\u524d\u7528 Qwen2.5-7B \u8dd1\u4e00\u4e2a&#8221;\u67e5\u6570\u636e\u2192\u753b\u56fe\u2192\u5199\u7ed3\u8bba&#8221;\u7684\u4e09\u6b65\u6d41\u7a0b\uff0c\u6210\u529f\u7387\u5927\u6982\u5728 60% \u5de6\u53f3\u3002\u6302\u4e0a Forge \u7684 proxy \u6a21\u5f0f\u540e\uff0c\u540c\u4e00\u4e2a\u6a21\u578b\u3001\u540c\u6837\u7684 prompt\uff0c\u8dd1 20 \u6b21\u6210\u529f\u4e86 18 \u6b21\u3002<\/p>\n<p>90% \u7684\u6210\u529f\u7387\u5bf9 7B \u6a21\u578b\u6765\u8bf4\u5df2\u7ecf\u76f8\u5f53\u80fd\u6253\u4e86\u3002<\/p>\n<p>\u5f53\u7136\u4e5f\u4e0d\u662f\u6ca1\u4ee3\u4ef7\u2014\u2014\u591a\u4e86 rescue \u91cd\u8bd5\u548c context compaction \u7684\u5f00\u9500\uff0c\u54cd\u5e94\u5ef6\u8fdf\u591a\u4e86\u5927\u6982 30%\u3002\u4f46\u5bf9\u4e8e\u79bb\u7ebf\u4efb\u52a1\u6216\u8005\u5bf9\u5ef6\u8fdf\u4e0d\u654f\u611f\u7684\u573a\u666f\uff0c\u8fd9\u70b9\u4ee3\u4ef7\u5b8c\u5168\u53ef\u4ee5\u63a5\u53d7\u3002<\/p>\n<p>## \u51e0\u70b9\u5410\u69fd<\/p>\n<p>1. **\u6587\u6863\u504f\u5de5\u7a0b\u5316**\u3002README \u5199\u5f97\u633a\u597d\uff0c\u4f46\u5177\u4f53\u914d\u7f6e\u9879\u7684\u8bf4\u660e\u4e0d\u591f\u7ec6\uff0c\u6709\u51e0\u4e2a\u53c2\u6570\u5f97\u7ffb\u6e90\u7801\u624d\u80fd\u641e\u660e\u767d\u3002<br \/>\n2. **Python 3.12+ \u9650\u5b9a**\u3002\u5982\u679c\u4f60\u8fd8\u5728\u7528 3.10\/3.11\uff0c\u5f97\u5148\u5347\u7ea7\u73af\u5883\u3002<br \/>\n3. **\u4ee3\u7406\u6a21\u5f0f\u8fd8\u4e0d\u652f\u6301\u6d41\u5f0f\u8f93\u51fa**\u3002\u8fd9\u5bf9\u4e8e\u804a\u5929\u573a\u666f\u6709\u70b9\u4f24\uff0c\u671f\u5f85\u4f5c\u8005\u540e\u7eed\u52a0\u4e0a\u3002<\/p>\n<p>## \u603b\u7ed3<\/p>\n<p>Forge \u8ba9\u6211\u91cd\u65b0\u601d\u8003\u4e86\u4e00\u4e2a\u4e8b\u60c5\uff1a\u5c0f\u6a21\u578b\u7684\u8fb9\u754c\u5230\u5e95\u5728\u54ea\uff1f<\/p>\n<p>\u4e4b\u524d\u5927\u5bb6\u7684\u5171\u8bc6\u662f&#8221;\u672c\u5730\u6a21\u578b\u5e72\u4e0d\u4e86\u590d\u6742 agent \u6d3b\uff0c\u5f97\u4e0a\u4e91&#8221;\u3002Forge \u7528\u5de5\u7a0b\u624b\u6bb5\u8bc1\u660e\u2014\u2014**\u95ee\u9898\u4e0d\u4e00\u5b9a\u51fa\u5728\u6a21\u578b\u8111\u5b50\u4e0d\u591f\uff0c\u800c\u662f\u7f3a\u4e86\u4e00\u5957\u597d\u7684\u7ba1\u7406\u6846\u67b6\u3002**<\/p>\n<p>\u5982\u679c\u4f60\u624b\u5934\u6709\u672c\u5730\u6a21\u578b\u60f3\u505a agent \u65b9\u5411\uff0cForge \u503c\u5f97\u82b1\u4e00\u4e2a\u4e0b\u5348\u8bd5\u8bd5\u3002\u5b83\u662f MIT \u534f\u8bae\uff0c\u968f\u4fbf\u62ff\u53bb\u6539\u3002GitHub \u5730\u5740\uff1a[antoinezambelli\/forge](https:\/\/github.com\/antoinezambelli\/forge)<\/p>\n<p>&#8212;<\/p>\n<p>*\u53e6\u5916\u4eca\u5929 Google I\/O \u53d1\u5e03\u4e86 Gemini 3.5 Flash\uff0c\u628a&#8221;\u884c\u52a8\u80fd\u529b&#8221;\u76f4\u63a5\u505a\u5230\u6a21\u578b\u5c42\u4e86\u3002Mistral \u4e5f\u6536\u8d2d\u4e86 Emmi AI \u5e03\u5c40\u5de5\u4e1a\u5236\u9020\u573a\u666f\u3002\u672c\u5730\u5c0f\u6a21\u578b\u548c\u4e91\u7aef\u5927\u6a21\u578b\u4e24\u6761\u8def\u7ebf\u90fd\u5728\u52a0\u901f\u2014\u2014\u8fd9\u4e00\u5e74\u7684 AI \u5de5\u7a0b\u5316\u8fdb\u7a0b\u6bd4\u8fc7\u53bb\u4e09\u5e74\u90fd\u5feb\u3002*<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u524d\u4e24\u5929\u901b Hacker News\uff0c\u770b\u5230\u4e00\u4e2a\u9879\u76ee\u8ba9\u6211\u773c\u524d\u4e00\u4eae\u2014\u2014**Forge**\uff0c\u4e00\u4e2a\u7ed9\u672c\u5730\u5c0f\u6a21\u578b\u52a0\u62a4\u680f\uff08Gu&hellip; <a href=\"https:\/\/zoran.090702.xyz\/?p=174\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb <span class=\"screen-reader-text\">\u628a8B\u5c0f\u6a21\u578b\u4ece53%\u62c9\u523099%\uff1aForge \u5f00\u6e90\u9879\u76ee\u5b9e\u6218\u89e3\u6790<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[],"class_list":["post-174","post","type-post","status-publish","format-standard","hentry","category-ai-frontier"],"_links":{"self":[{"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=\/wp\/v2\/posts\/174","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=174"}],"version-history":[{"count":1,"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=\/wp\/v2\/posts\/174\/revisions"}],"predecessor-version":[{"id":175,"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=\/wp\/v2\/posts\/174\/revisions\/175"}],"wp:attachment":[{"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=174"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=174"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/zoran.090702.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=174"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}