unhardcoded
开源 · AI 模型的策略路由层

构建 AI 系统，无需硬编码模型选择。

每次请求附带一条策略。unhardcoded 过滤实时模型目录，选出满足规则的成本最低的候选模型，通过你的提供商密钥运行，并返回决策追踪记录。

request
Prompt + policy
tools · intel ge 0.5 · cheapest
→
candidateprice_outintelverdict
deepseek-v4-flash$0.400.465未达下限
minimax-m2.7$0.500.496未达下限
deepseek-v4-pro$1.500.515胜出
glm-5.1$2.000.514通过
gpt-5.5$10.000.602通过
→
trace
deepseek-v4-pro
cheapest passing · 412 ms
fp 301140696-1054914287
路由单次调用
将 SDK 指向托管地址，附加策略，读取追踪记录。
快速入门 →
运行工作流
将路由调用组合成 DAG：分类、扇出、评判、合并。
查看模式 →
读取追踪记录
每次决策都留下凭证：谁被考虑过，谁胜出，以及原因。
检查决策 →
核心模型

模型路由循环

一条策略决定一次调用。相同输入、相同目录、相同决策。每次如此，且有书面记录。

Request
你的调用到达兼容 OpenAI 的端点，携带一条策略。
Policy sigma-pol/v2
一个小型项(term)：filter、rank、select、mutate、fallback。
Catalog
包含价格、基准分数和能力的（提供商、模型）候选模型实时集合。
Filter
剔除不满足规则的候选模型，不存在静默降级。
Rank
对通过筛选的候选模型评分：按价格、能力或速度排序，由你决定。
Select
取排名靠前的模型，或使用 top_k 回退级联。
Run
通过你的提供商密钥运行推理，出错时自动回退。
Trace
记录所有候选模型、胜出者及指纹(fingerprint)的凭证。

深入了解策略工作原理 →

组合路由调用

工作流模式

工作流是由路由步骤组成的有界无环图。每个节点携带自己的策略并独立路由；整个图生成一条统一的追踪记录。这正是 unhardcoded 从"带规则的路由"演变为一个完整系统的地方。

支持工单linear · guard

低成本分类，按质量下限起草回复，最后由一个强力 no-log 守卫在发送前决定是否拒绝。

关键点：最后一步是由策略强制执行的门控，可以中止发送，而不仅仅是希望它有效。

工单
→
triageclassify · extract iddeepseek-v4-flash
→
draftreply · quality floorgemini-3.1-pro-preview
→
guardbrand · PII · no-loggpt-5.5 · can abort
→
回复
查看 flow_ir
flow.support-ticket.json
["flow", {
  "u": {"kind": "input"},
  "t": {"kind": "llm", "system": "Classify the ticket and extract the account id as JSON.",
    "policy": ["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]], ["has_cap", "supports_json_mode"]],
      ["neg", ["normalize", ["field", "price_out"]]], ["argmax"], ["id"], ["always", {"action": "next_candidate"}]],
    "inputs": ["u"]},
  "d": {"kind": "llm", "system": "Write a reply using the ticket and the triage.",
    "policy": ["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]], ["cmp", "bench_intelligence", "ge", 0.55]],
      ["neg", ["normalize", ["field", "price_out"]]], ["argmax"], ["id"], ["always", {"action": "next_candidate"}]],
    "inputs": ["u", "t"], "template": "Ticket:\n$1\n\nTriage:\n$2"},
  "g": {"kind": "llm", "system": "Check brand voice, PII, refund limits. Refuse if any fail.",
    "policy": ["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]], ["is", "no_log"]],
      ["field", "bench_intelligence"], ["argmax"], ["id"], ["always", {"action": "next_candidate"}]],
    "inputs": ["d"]},
  "out": {"kind": "output", "inputs": ["g"]}
}]
起草 → 批评 → 修改linear · refine

先用低成本模型生成初稿，再由强力批评者列出缺陷，最后重写并修复每个问题。

关键点：在预算范围内实现质量跃升——大部分令牌消耗在低成本模型上。

问题
→
draftcheapestdeepseek-v4-flash
→
critiquelist the flawsgpt-5.5
→
revisefix every point · fan-ingpt-5.5
→
答案
查看 flow_ir
flow.draft-critique-revise.json
["flow", {
  "u": {"kind": "input"},
  "d": {"kind": "llm", "system": "Draft an answer.",
    "policy": ["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]]],
      ["neg", ["normalize", ["field", "price_out"]]], ["argmax"], ["id"], ["always", {"action": "next_candidate"}]],
    "inputs": ["u"]},
  "c": {"kind": "llm", "system": "Critique the draft: list concrete flaws and gaps.",
    "policy": ["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]]],
      ["field", "bench_intelligence"], ["argmax"], ["id"], ["always", {"action": "next_candidate"}]],
    "inputs": ["d"]},
  "r": {"kind": "llm", "system": "Rewrite the answer, fixing every point in the critique.",
    "policy": ["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]]],
      ["field", "bench_intelligence"], ["argmax"], ["id"], ["always", {"action": "next_candidate"}]],
    "inputs": ["u", "d", "c"], "template": "Question:\n$1\n\nDraft:\n$2\n\nCritique:\n$3"},
  "out": {"kind": "output", "inputs": ["r"]}
}]
N 选一 + 评判ensemble

同一强力策略下进行 N 次种子采样，然后由评判者从中择优选出胜出结果。

关键点：sample 以确定性方式在排名靠前的候选模型中分散采样——可复现的多样性，而非随机。

提示词
→
draft Asample · T=0.5gpt-5.5
draft Bsample · T=0.5gpt-5.4
draft Csample · T=0.5gemini-3.1-pro-preview
→
judgerank · pick winnergpt-5.5
→
答案
查看 flow_ir
flow.best-of-n.json
["flow", {
  "u": {"kind": "input"},
  "n1": {"kind": "llm", "system": "Answer the question.",
    "policy": ["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]]],
      ["field", "bench_intelligence"], ["sample", 0.5], ["id"], ["always", {"action": "next_candidate"}]],
    "inputs": ["u"]},
  // n2, n3: the same policy, two more seeded draws
  "j": {"kind": "llm", "system": "Pick the single best candidate; return it verbatim.",
    "policy": ["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]]],
      ["field", "bench_intelligence"], ["argmax"], ["id"], ["always", {"action": "next_candidate"}]],
    "inputs": ["n1", "n2", "n3"], "template": "A:\n$1\n\nB:\n$2\n\nC:\n$3"},
  "out": {"kind": "output", "inputs": ["j"]}
}]
模型评审团fan-out · fuse

三个按族系固定的模型并行起草，第四个模型将它们合成为一个更优的单一答案。

关键点：family_eq 锁定了确切的模型系列，因此评审团可复现——而不是"今天最便宜的那个"。

问题
→
draft afamily_eqgemini-3.1-pro-preview
draft bfamily_eqclaude-opus-4-8
draft cfamily_eqdeepseek-v4-flash
→
fusesynthesize · fan-ingpt-5.5
→
答案
查看 flow_ir
flow.panel.json
["flow", {
  "u": {"kind": "input"},
  "a": {"kind": "llm", "system": "Draft an answer.",
    "policy": ["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]], ["family_eq", "gemini-3.1-pro-preview"]],
      ["zero"], ["argmax"], ["id"], ["always", {"action": "next_candidate"}]], "inputs": ["u"]},
  // b: family_eq "claude-opus-4-8"  ·  c: family_eq "deepseek-v4-flash"
  "f": {"kind": "llm", "system": "Synthesize the single best answer from the drafts.",
    "policy": ["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]]],
      ["field", "bench_intelligence"], ["argmax"], ["id"], ["always", {"action": "next_candidate"}]],
    "inputs": ["a", "b", "c"]},
  "out": {"kind": "output", "inputs": ["f"]}
}]
专家分工fan-out · merge

推理器和编码器在不同过滤条件下并行运行，之后由合并步骤整合。

关键点：每个分支为其任务（推理 vs 编码）选择最合适的模型，而非让一个模型包揽所有工作。

问题
→
reasonis cap_reasoninggpt-5.5
codecoding top-5gpt-5.4
→
mergeone answer · fan-ingemini-3.1-pro-preview
→
答案
查看 flow_ir
flow.specialist-split.json
["flow", {
  "u": {"kind": "input"},
  "rz": {"kind": "llm", "system": "Reason through the problem step by step.",
    "policy": ["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]], ["is", "cap_reasoning"]],
      ["field", "bench_intelligence"], ["argmax"], ["id"], ["always", {"action": "next_candidate"}]], "inputs": ["u"]},
  "cd": {"kind": "llm", "system": "Produce any code the problem needs.",
    "policy": ["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]], ["cmp", "bench_coding_rank", "le", 5]],
      ["field", "bench_coding"], ["argmax"], ["id"], ["always", {"action": "next_candidate"}]], "inputs": ["u"]},
  "m": {"kind": "llm", "system": "Merge the reasoning and the code into one answer.",
    "policy": ["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]]],
      ["field", "bench_intelligence"], ["argmax"], ["id"], ["always", {"action": "next_candidate"}]],
    "inputs": ["rz", "cd"], "template": "Reasoning:\n$1\n\nCode:\n$2"},
  "out": {"kind": "output", "inputs": ["m"]}
}]

完整工作流指南：节点类型、图限制、模板 →

发送初始请求

快速入门

unhardcoded 兼容 OpenAI 接口。相比普通调用，仅需三处改动。

  1. 将 SDK 指向托管地址。
    修改 baseURL；SDK 的其余部分保持不变。
    client.ts
    const client = new OpenAI({
  baseURL: "https://<your-host>/v1",
  apiKey: process.env.UNHARDCODED_KEY,
});
  2. 在调用中附加策略。
    在后端构建 policy_ir，与 messages 一起发送。路由来自策略，因此 model 只是追踪记录标签。
    route.ts
    const res = await client.chat.completions.create({
  model: "policy:support",
  policy_ir: ["policy",
    ["and", ["meets_req"], ["not", ["is", "disabled"]],
           ["cmp", "bench_intelligence", "ge", 0.5]],   // filter
    ["neg", ["normalize", ["field", "price_out"]]],          // rank: cheapest
    ["argmax"], ["id"], ["always", {"action": "next_candidate"}]],
  messages,
});
  3. 读取追踪记录。
    响应中包含决策内容：所选模型、经过排序和拒绝的候选模型，以及策略指纹(fingerprint)。
    response · trace (illustrative)
    {
  "chosen": { "model_family": "deepseek-v4-pro", "price_out": 1.5 },
  "trace": {
    "policy_fingerprint": "301140696-1054914287",
    "rejected": [{ "model_family": "deepseek-v4-flash", "reason": "cmp bench_intelligence ge 0.5" }],
    "total_latency_ms": 425
  }
}

完整快速入门：认证、演练运行、可运行示例 →

复制一个起点

策略预设

以卡片形式呈现的常用路由模式。阅读规则，复制策略，调整下限和上限。

低成本适当质量
在不低于质量下限的前提下降低成本。
filtertools-met · bench_intelligence ge 0.5
rankcheapest price_out
fallback下一个通过的候选模型
查看 JSON
cheapest-decent.json
["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]],
         ["cmp", "bench_intelligence", "ge", 0.5]],
  ["neg", ["normalize", ["field", "price_out"]]],
  ["argmax"], ["id"], ["always", {"action": "next_candidate"}]]
智能均衡
无强烈偏好：综合权衡能力与价格。
filtertools-met · not disabled
rank0.6 intelligence + 0.4 cheap
fallback下一个通过的候选模型
查看 JSON
smart-balance.json
["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]]],
  ["add",
    ["scale", 0.6, ["normalize", ["field", "bench_intelligence"]]],
    ["scale", 0.4, ["neg", ["normalize", ["field", "price_out"]]]]],
  ["argmax"], ["id"], ["always", {"action": "next_candidate"}]]
优先智能能力
关键任务，无成本上限：选择能力更胜一筹的模型。
filtertools-met · not disabled
rankhighest bench_intelligence
fallback下一个通过的候选模型
查看 JSON
best-intelligence.json
["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]]],
  ["field", "bench_intelligence"],
  ["argmax"], ["id"], ["always", {"action": "next_candidate"}]]
仅限推理模型
任务需要具备推理能力的模型。
filter+ is cap_reasoning
rankhighest intelligence
fallback下一个通过的候选模型
查看 JSON
reasoning-only.json
["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]], ["is", "cap_reasoning"]],
  ["field", "bench_intelligence"],
  ["argmax"], ["id"], ["always", {"action": "next_candidate"}]]
视觉 · 低成本
图像输入：能够处理图像的成本最低模型。
filter+ is in_image
rankcheapest price_out
fallback下一个通过的候选模型
查看 JSON
vision-cheapest.json
["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]], ["is", "in_image"]],
  ["neg", ["normalize", ["field", "price_out"]]],
  ["argmax"], ["id"], ["always", {"action": "next_candidate"}]]
长上下文 RAG
需要大上下文窗口时，选择能满足要求的成本最低模型。
filter+ context ge 200000
rankcheapest price_out
fallback下一个通过的候选模型
查看 JSON
long-context-rag.json
["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]], ["cmp", "context", "ge", 200000]],
  ["neg", ["normalize", ["field", "price_out"]]],
  ["argmax"], ["id"], ["always", {"action": "next_candidate"}]]
智能体调度
为智能体循环选择工具调用能力出色的模型。
filter+ tools · bench_agentic_rank le 5
rankhighest bench_agentic
fallback下一个通过的候选模型
查看 JSON
agentic-fleet.json
["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]],
         ["has_cap", "supports_tools"], ["cmp", "bench_agentic_rank", "le", 5]],
  ["field", "bench_agentic"],
  ["argmax"], ["id"], ["always", {"action": "next_candidate"}]]
私有 / 合规
仅限 TEE 且禁止日志记录：能力优先。
filter+ is has_tee · is no_log
rankhighest intelligence
fallback下一个通过的候选模型
查看 JSON
private-compliant.json
["policy", ["and", ["meets_req"], ["not", ["is", "disabled"]],
         ["is", "has_tee"], ["is", "no_log"]],
  ["field", "bench_intelligence"],
  ["argmax"], ["id"], ["always", {"action": "next_candidate"}]]

所有预设及实用配方 →

信任凭证

每次决策都留下凭证

追踪记录是模型选择过程的结构化、可重放记录：哪些模型被考虑过，每个模型为何通过或被拒绝，执行了什么，以及如何复现。

decision_trace
deepseek-v4-flashrejected · cmp bench_intelligence ge 0.5
minimax-m2.7rejected · cmp bench_intelligence ge 0.5
deepseek-v4-proranked #1 · chosen
glm-5.1ranked #2 · cascade
gpt-5.5ranked #3 · cascade
deepseek-v4-prodecision_path · attempted · 412 ms
policy_fingerprint 301140696-1054914287 · sigma-pol/v2
rejected[]
所有被过滤掉的候选模型，以及导致其被拒绝的确切规则。
ranked[]
按选择顺序排列的通过候选模型；排名靠前的是选中的模型，其余为回退级联。
decision_path[]
尝试日志；每次回退跳转都记录在此。
policy_fingerprint
归一化策略的稳定 ID；标识符由托管侧的 sha256 计算。工作流还包含 flow_fingerprint。

重放与调试的工作原理 →

需要细节时

参考文档

完整规范，独立于阅读路径之外。

API 参考
端点、认证、请求结构
policy_ir
语法与五个插槽
flow_ir
节点类型、图限制、模板
算子 & 字段
签名与目录词汇表
追踪记录 & 重放
追踪模式、合规性
策略
概念、预设、配方
自托管
托管服务、目录、提供商密钥
故障排查
错误、无候选模型、回退

GitHub 源码：unhardcoded（托管服务）· unhardcoded-engine（策略引擎）