Jev 是什么
怎么用，怎么跟大模型配合

# 第 62–66 行 t = time.perf_counter() r = jev.decide({"ticket": ticket}, Choice("Which queue...?", QUEUES)) show("", r, time.perf_counter() - t)

# 第 74–79 行 t = time.perf_counter() r = jev.decide({"ticket": ticket}, Boolean("Is the customer...?", ...)) show("", r, time.perf_counter() - t)

# 第 81–85 行 t = time.perf_counter() r = jev.decide({"ticket": ticket}, URGENCY) show("", r, time.perf_counter() - t)

# TICKETS[3] 是这一句： "The dashboard takes about eight seconds to load..." # 但 QUEUES 里只有三个部门： Option("access", ...) Option("billing", ...) Option("sales", ...) # 没有「技术支持」这个选项

from fastjev import Choice, Option QUEUES = ( Option("access", "Login and account access."), Option("billing", "Billing, payments, and refunds."), Option("sales", "Pricing and new contracts."), ) r = jev.decide( {"ticket": "I was charged twice, need a refund."}, Choice("Which queue should handle this?", QUEUES), ) r.value # "billing" r.probabilities # {"billing": 0.9997, ...}

from fastjev import Boolean r = jev.decide( {"ticket": ticket}, Boolean( "Is the customer asking for money back?", true_description="They ask for a refund.", false_description="They do not ask for money.", ), ) r.value # True r.probabilities # {True: 0.9998, False: 0.0002}

from fastjev import Level, Score URGENCY = Score("How urgently does this need a human?", ( Level(1.0, "No urgency; can wait."), Level(2.0, "Low; within a few days."), Level(3.0, "Normal; within one business day."), Level(4.0, "High; needs attention today."), Level(5.0, "Critical; service is down."), )) r = jev.decide({"ticket": ticket}, URGENCY) r.value # 4.13 ← 概率加权后的数值，可以是小数 r.selected # 5.0 ← 概率最高的那一档

results = jev.decide_many( {"ticket": ticket}, { "queue": Choice("Which queue handles this?", QUEUES), "refund": Boolean("Do they want money back?"), "urgency": URGENCY, }, ) results["queue"].value # "billing" results["refund"].value # True results["urgency"].value # 4.13

RULES = { "spam": Boolean("Is this spam or an ad?"), "abuse": Boolean("Does this attack a person?"), "unsafe": Boolean("Is this dangerous medical advice?"), } flags = jev.decide_many({"comment": c}, RULES) hit = [n for n, d in flags.items() if d.value] if hit: send_to_review(c, hit) # 哪几条踩了就一起报

docs = search(query) # 先粗筛出 10 条 cands = {f"d{i}": d for i, d in enumerate(docs)} opts = tuple(Option(k, v.text) for k, v in cands.items()) r = jev.decide( {"query": query}, Choice("Which passage best answers it?", opts), ) best = cands[r.value] # r.value 是 "d3" 这种 id

r = jev.decide( {"q": user_q, "a": llm_output}, Boolean( "Does the answer address what was asked?", true_description="It responds to the question.", false_description="It dodges the question.", ), ) if not r.value or r.probabilities[False] > 0.2: escalate(llm_output) # 拿不准就别直接给用户

DIFFICULTY = Choice("How hard is this request?", ( Option("simple", "Short and factual; one step."), Option("hard", "Multi-step reasoning or math."), )) r = jev.decide({"prompt": prompt}, DIFFICULTY) model = "small" if r.value == "simple" else "large" answer = call_llm(model, prompt) # 难才用贵的

RELEVANCE = Score("How relevant is this to the query?", ( Level(0.0, "Unrelated to the query."), Level(1.0, "Same topic, but does not answer it."), Level(2.0, "Partially answers it."), Level(3.0, "Fully answers it."), )) def score_of(item): state = {"query": q, "text": item} return jev.decide(state, RELEVANCE).value ranked = sorted(items, key=score_of, reverse=True)

ROUTE = Choice("Which model can handle this?", ( Option("small", "Rename, format, simple lookup."), Option("large", "Design, debugging, multi-file change."), )) r = jev.decide({"task": task}, ROUTE) model = "haiku" if r.value == "small" else "sonnet" answer = call_llm(model, task) # 你的大模型调用

files = list_files() # 代码先把候选列出来 cands = {f"f{i}": p for i, p in enumerate(files)} r = jev.decide( {"task": task}, Choice("Which file must be read to do this?", tuple(Option(k, v) for k, v in cands.items())), ) read(cands[r.value]) # 只把挑中的那个喂给大模型

r = jev.decide( {"task": task, "diff": diff}, Boolean( "Does this diff do what was asked?", true_description="It implements the request.", false_description="It misses it or edits other code.", ), ) if not r.value or r.probabilities[False] > 0.2: retry(task) # 拿不准就重来，别直接交

DANGER = Boolean( "Could this command destroy data?", true_description="Deletes, overwrites, or hits production.", false_description="Read-only, or scoped to this repo.", ) r = jev.decide({"command": cmd}, DANGER) if r.value or r.probabilities[True] > 0.05: ask_human(cmd) # 宁可多问一句

from fastjev import FastJev # 模型只加载一次，整个 agent 生命周期共用 with FastJev.from_pretrained(MODEL_DIR) as jev: r = jev.decide({"task": task}, ROUTE) model = "haiku" if r.value == "small" else "sonnet"

# 终端里起服务 fastjev-serve \ --model Qwen/Qwen3.5-4B \ --revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a \ --served-model fastjev-qwen3.5-4b \ --served-model-description 'fastjev on Qwen3.5-4B' \ --served-model-release-date 2026-09-18 # 然后任何能发 HTTP 的节点都能调它 POST http://127.0.0.1:8000/v1/systemone { "model": "fastjev-qwen3.5-4b", "state": "帮我重命名这个函数", "questions": { "danger": {"type": "noul", "instructions": "Is this destructive?"}, "route": {"type": "choice", "instructions": "Which model fits?", "criteria": { "small": "Rename, format, lookup.", "large": "Design or debugging." }} } }

{ "model": "fastjev-qwen3.5-4b", "answers": { "danger": {"type": "noul", "noul": 0.03}, "route": { "type": "choice", "choice": "small", "probabilities": {"small": 0.97, "large": 0.03}, "confidence": 0.88 } }, "usage": {"input_tokens": 214, "output_tokens": 0} }