← Frameworks

HaystackHaystack

deepset

开源Open source Apache2.0Apache2.0 RAGRAG 检索管线Retrieval pipeline 生产级Production-grade partial lifecycle: active 核验Checked 2026-10-30

Haystack 是「检索质量工程」那一派的 RAG 框架（Apache-2.0，deepset）：索引/查询管线分离、 hybrid 检索与 reranker、评估环——生产味最浓的一条 RAG 线。2.x 管线 API 大改，旧教程要警惕。

Haystack is the RAG framework of the 'retrieval-quality engineering' school (Apache-2.0, deepset): separated indexing/query pipelines, hybrid retrieval with rerankers, an evaluation loop — the most production-flavored RAG line. The 2.x pipeline API was rewritten; beware old tutorials.

适合与不适合When to use it, when not to

适合：生产 RAG（要检索质量与可审计组件图）、自建存储的数据主权场景。不适合：薄聊天壳；深度 agent 编排（LangGraph 更专）。

Fits: production RAG (retrieval quality and auditable component graphs), data sovereignty on self-chosen storage. Not a fit: thin chat shells; deep agent orchestration (LangGraph is more specialized).

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　deepset GmbH（德国柏林）。deepset GmbH (Berlin, Germany).
国内使用Direct access: 大陆可直连direct OK　haystack.deepset.ai 直连实测 200（2026-10-21）。haystack.deepset.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: 各家生成模型与嵌入模型经集成接入（OpenAI/Anthropic/本地 HF 等，以官方集成目录为准，未逐条清点）；RAG 检索栈（dense/sparse/hybrid）是它的看家面。
Generation and embedding models plug in via integrations (OpenAI/Anthropic/local HF and the like, per the official integration catalog, not itemized one by one); the RAG retrieval stack (dense/sparse/hybrid) is its home turf.
运行位置Where it runs: Python 库（pip 装进你的进程）；GitHub 26,880★，最近推送 2026-10-30（API 核验）。管线（pipeline）抽象是主轴：组件连成图，跑批或跑服务都在你的进程里。
A Python library (pip-installed into your process); GitHub 26,880★, last push 2026-10-30 (API verified). The pipeline abstraction is the main axis: components wired into a graph, batch or serving — it all runs inside your process.
本地文件Local files: 文档索引与检索数据在你选的存储（文件/向量库/文档库）——数据不出你侧； deepset Cloud 托管线是另一产品面。
Document indexes and retrieval data live in the storage you choose (files/vector stores/document stores) — data never leaves your side; the deepset Cloud hosted line is a separate product surface.
关机后的任务Tasks after shutdown: 库无常驻语义；索引管道可离线批跑，查询服务由你部署。 2.x 把 1.x 的 node 体系重写成 component/pipeline 模型，升级是大改而非平滑迁移。
A library has no resident semantics; indexing pipelines can run offline in batches, and the query service is deployed by you. 2.x rewrote 1.x's node system into the component/pipeline model — upgrading is a major rewrite, not a smooth migration.
工具与扩展Tools & extensions: agent 层（2.x 起）含工具调用循环；MCP 支持状态未逐条核验。
The agent layer (since 2.x) includes a tool-calling loop; MCP support status not itemized.
上下文与记忆Context & memory: 会话记忆组件存在（chat message store 线）；跨会话记忆要自己接存储——组件给模式不给托管。
A conversation-memory component exists (the chat message store line); cross-session memory requires wiring your own storage — the component gives the pattern, not hosting.
权限与限制Permissions & limits: 无账号体系（库形态）；权限面=你接的各家 key 与存储凭据。deepset Cloud 上传面是另一条线的口径。
No account system (library form); the permissions surface = the various vendors' keys and storage credentials you wire in. The deepset Cloud upload surface follows a different line's terms.
适合什么任务Best for: 适合：生产级 RAG 管线（检索质量工程：hybrid/reranker/评估环）、要可审计的组件图。不适合：只要一个薄聊天壳（过重）；agent 深度编排（LangGraph 线更专）。
Fits: production-grade RAG pipelines (retrieval-quality engineering: hybrid/reranker/evaluation loops), component graphs you can audit. Does not fit: just wanting a thin chat shell (too heavy); deep agent orchestration (the LangGraph line is more specialized).

头号误解Top misconception

把 Haystack 当「demo 级 RAG 脚本」——它的定位是生产管线（索引/查询分离、评估环）Treating Haystack as a demo-grade RAG script — its positioning is production pipelines (index/query separation, evaluation loops)
把 deepset Cloud 与开源库混谈——托管线的数据上传口径是另一产品条款Blurring deepset Cloud with the open-source library — the hosted line's data-upload terms are a separate product's clauses
1.x→2.x 的管线 API 大改——旧教程代码在新版跑不通，看版本标记The 1.x→2.x pipeline API rewrite — old-tutorial code breaks on the new version; check version markers

未知项清单Known unknowns

集成全目录（未逐条清点）
agent 层当前完成度与 MCP 支持（未逐条核验）
deepset Cloud 的数据处理条款细节（另一产品面，未逐条核验）

The full integration catalog (not itemized one by one)
Current completeness of the agent layer and MCP support (not itemized)
Details of deepset Cloud's data-processing terms (a separate product surface, not itemized)

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	deepset-ai/haystack · 仓库（Apache-2.0；26,880★，最近推送 2026-10-30，GitHub API 核验）deepset-ai/haystack · repository (Apache-2.0; 26,880★, last push 2026-10-30, GitHub API verified)	https://github.com/deepset-ai/haystack
docs	Haystack · 官方文档Haystack · official docs	https://haystack.deepset.ai
changelog	deepset-ai/haystack · Releases（changelog）deepset-ai/haystack · Releases (changelog)	https://github.com/deepset-ai/haystack/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.