← Frameworks

DSPyDSPy

Stanford NLP / Databricks

开源Open source MITMIT 提示优化Prompt optimisation 编程式LMProgrammatic LM 学术出身Academic origins partial lifecycle: active 核验Checked 2026-10-30

DSPy 把「调提示词」变成编程问题（MIT，Stanford NLP/Databricks）：签名声明输入输出，优化器把提示当参数训练。用它的分水岭一句话：有没有评估指标——有指标它是杠杆，没指标它无从下手。

DSPy turns 'tuning prompts' into a programming problem (MIT, Stanford NLP/Databricks): signatures declare inputs/outputs, optimizers train prompts as parameters. The dividing line in one sentence: do you have an evaluation metric — with one it is a lever, without one it has nothing to optimize.

适合与不适合When to use it, when not to

适合：可复现的提示优化流程、多模型可换管线、有 metric 的成熟任务。不适合：开箱 agent 运行时；无指标的探索期任务。

Fits: reproducible prompt-optimization flows, multi-model swappable pipelines, mature tasks with metrics. Not a fit: out-of-the-box agent runtimes; exploratory tasks without metrics.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-21

出品属地Origin: 国外出品foreign-made　Stanford NLP 起源，现与 Databricks 有官方投入线。Stanford NLP origin, with official Databricks investment lines.
国内使用Direct access: 大陆可直连direct OK　dspy.ai 直连实测 200（2026-10-21）。dspy.ai measured 200 direct (2026-10-21).
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-21 检索）；开源项目未面向大陆公众提供服务，备案不适用。not in the CAC public registry (searched 2026-10-21); the open-source project does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 各家 LM 通过 dspy.LM 统一接入（OpenAI/Anthropic/Gemini/本地 vLLM 等，以官方 docs 的 LM 集成线为准，未逐条清点）；它的差异点是把提示词当可编程参数—— 签名（signature）声明输入输出，优化器自动调提示与权重示例。
Each vendor's LM connects uniformly through dspy.LM (OpenAI/Anthropic/Gemini/local vLLM etc., following the official docs' LM-integration line, not enumerated item by item); its differentiator is treating prompts as programmable parameters — signatures declare inputs and outputs, and optimizers automatically tune prompts and few-shot examples.
运行位置Where it runs: Python 库（pip 装进你的进程）；GitHub 38,850★，最近推送 2026-10-29（API 核验）。学术出身（Stanford NLP），工业侧 Databricks 投入（官方叙事口径）。
A Python library (pip-installed into your process); GitHub 38,850★, last push 2026-10-29 (API verified). Academic origin (Stanford NLP), with industrial-side Databricks investment (per the official narrative).
本地文件Local files: 库形态跑在你进程里；优化产物（program 状态/示例）落你指定路径——数据不出你侧。
The library form runs inside your process; optimization artifacts (program state/examples) land at paths you specify — data never leaves your side.
关机后的任务Tasks after shutdown: 库无常驻语义；优化（teleprompting）是离线批任务，跑完即止。
A library has no resident semantics; optimization (teleprompting) is an offline batch task that ends when it finishes.
工具与扩展Tools & extensions: ReAct 类 agent 模块与工具调用在官方模块族内；MCP 支持状态未核验到官方主线声明。
ReAct-style agent modules and tool calling are within the official module family; MCP support status has no verified official mainline statement.
上下文与记忆Context & memory: 状态管理跟着你的 program 走；跨会话记忆不是它的叙事——它的「记忆」是优化出来的示例与权重。
State management follows your program; cross-session memory is not part of its narrative — its "memory" is the optimized examples and weights.
权限与限制Permissions & limits: 无账号体系；权限面=你接的各家 API key。优化过程会大量调模型（token 成本面），预算控制参数在官方文档线，细节未逐条核验。
No account system; the permission surface is whatever vendors' API keys you connect. The optimization process calls models heavily (a token-cost surface); budget-control parameters are in the official docs line, details not verified item by item.
适合什么任务Best for: 适合：把「调提示词」变成可复现的优化流程（有指标、有训练集）、多模型可换的管线。不适合：要开箱 agent 运行时；小样本没指标时（优化器吃指标，没指标没意义）。
Fits: turning "prompt tuning" into a reproducible optimization process (with metrics, with a training set); pipelines where models are swappable. Does not fit: wanting an out-of-the-box agent runtime; small samples without metrics (optimizers consume metrics — without metrics there is nothing to optimize).

头号误解Top misconception

把 DSPy 当「又一个 prompt 模板库」——它的核心是优化器（把提示当参数训练），不用优化器等于白用Treating DSPy as 'yet another prompt template library' — the core is the optimizer (prompts trained as parameters); skipping it wastes the tool
没有评估指标就上优化器——优化器需要 metric，没指标时它无从优化Running optimizers without an evaluation metric — optimizers need a metric; without one there is nothing to optimize
优化期 token 成本不设上限——先设预算参数再跑（官方文档有控制面）Leaving optimizer token budgets uncapped — set budget parameters before running (the official docs have controls)

未知项清单Known unknowns

LM 集成全目录（未逐条清点）
MCP 支持状态（未核验到官方主线声明）
各优化器的样本量经验值（官方有指引，未逐条核验）

The full LM-integration catalog (not enumerated item by item)
MCP support status (no verified official mainline statement)
Sample-size rules of thumb per optimizer (the official docs have guidance; not verified item by item)

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-30，内容更新日content updated on 2026-10-21。

类型Type	名称Name	链接Link
repo	stanfordnlp/dspy · 仓库（MIT；38,850★，最近推送 2026-10-29，GitHub API 核验）stanfordnlp/dspy · repository (MIT; 38,850★, last push 2026-10-29, GitHub API verified)	https://github.com/stanfordnlp/dspy
docs	DSPy · 官方文档DSPy · official docs	https://dspy.ai
changelog	stanfordnlp/dspy · Releases（changelog）stanfordnlp/dspy · Releases (changelog)	https://github.com/stanfordnlp/dspy/releases

实测记录Measured runs

本档未跑集成与性能实测；结论只覆盖 API 与官方文档披露的能力面。

This site has not completed measured runs.

The test protocol is at protocols/framework/_protocol.md.