CocoIndex

智能体与MCP 11.8k 923 Python 上架 2026-11-08 访问官网 ↗ 在 GitHub 打开

CocoIndex 将代码库、会议记录、收件箱、Slack、PDF 和视频转化为实时、持续更新的上下文，供您的 AI 代理和 LLM 应用高效推理——仅需极少的增量处理。在 10 分钟内让您的生产级 ...

#Agent

项目信息

所有者: cocoindex-io
Star: 11.8k
Fork: 923
开源协议: —
主要语言: Python
发布时间: 2025-03-03
最近更新: 2026-10-28

项目更新

v1.0.262026-10-22
安全方面，所有 litellm 依赖（litellm、entityresolutionllm、all extras 及示例）已排除被投毒的 litellm 1.107.0，锁定版本维持 1.96.2。预览功能现在能在真实流水线上跑通且完全不落盘：支持 Postgres、Qdrant、LanceDB 等容器内的目标状态，不再一启动就中断，也不会写入 memo 或已提交状态，因此不会让下一次真实更新漏跑任务。连接器作者需改用 TargetActionSink.fromfn(fn, childhandler=...) / fromasyncfn(...)，旧的 fromfnwith_children 系列已废弃且不再支持预览。引擎新增 maxinflightcomponents 并发上限，父组件按需把额度借给子组件，宽扇出不超标、任意深度也不会死锁；coco.logicpropagatedto_callers(fn) 让回调的改动正确失效调用方缓存。 Rust SDK 修复了 memo 返回陈旧值、缓存命中误删子输出等问题；破坏性变更：neo4j/falkordb 目标函数改用 &ContextKey<Graph>。其余修复：`coco
v1.0.252026-10-06
What's Changed * feat(env): add Environment.close() to release the LMDB environment explicitly, so the same db path can be reopened without waiting for every Python reference to be collected (fixes EnvAlreadyOpened on free-threaded Python 3.14, where threads started inside a component kept the environment alive) by @georgeh0 in https://github.com/cocoindex-io/cocoindex/pull/2481 feat(memo): a memo state function can return recollectafter_run=True in MemoStateOutcome to have its state collected again after a re-run, for states that depend on what the function body read by @georgeh0 in https://github.com/cocoindex-io/cocoindex/pull/2474 Rust SDK ergonomics: #[cocoindex::function(memo, batching)] with per-item memoization, the contextkey! macro, and connector schemas derived from Rust row types (SchemaFields, TableSchema::from_row) with vector dimensions resolved at runtime from the embedder by @tomz-alt in https://github.com/cocoindex-io/cocoindex/pull/2287 * GPUPool rewritten in Rust; it can now acquire multiple whole GPUs at once, and no longer loses wake-ups across event loops by @YaxinCheng in https://github.com/cocoindex-io/cocoindex/pull/2277 * Performance for large target updates * Much less memory per declared target state: declared states are released right after pre-commit, and Python row values are held in encoded form (a 1.14 M-row component previously held ~1.33 GiB of row dicts) by @georgeh0 in https://github.com/cocoindex-io/cocoindex/pull/2469 fingerprintobject / memo fingerprints compute plain data natively (~9× faster per row); existing fingerprints are unchanged by @georgeh0 in https://github.com/cocoindex-io/cocoindex/pull/2470 Target handlers can set tracksvalue_fingerprint = True to skip reconcile() for unchanged values, so an update where most rows are unchanged runs no per-row Python at commit by @georgeh0 in https://github.com/cocoindex-io/cocoindex/pull/2472 * Postgres: upserts go through INSERT ... SELECT FROM unnest(arrays) for tables with builtin scalar/array columns by @georgeh0 in https://github.com/cocoindex-io/cocoindex/pull/2471 * Postgres: multi-column-key deletes use IN (VALUES ...) instead of an OR chain whose planning time grew quadratically by @georgeh0 in https://github.com/cocoindex-io/cocoindex/pull/2458 * fix(core): when one component's write fails (e.g. a target connector's reconcile() raises), only that component fails, not every component batched into the same transaction (including other apps sharing the Environment) by @georgeh0 in https://github.com/cocoindex-io/cocoindex/pull/2451 fix(core): scoped coco.exceptionhandlers now receive failures of mounteach items from a LiveMap (and autorefresh items) by @georgeh0 in https://github.com/cocoindex-io/cocoindex/pull/2478 * fix(core): preview no longer reports actions more than once when its write batch re-runs after the LMDB map grows by @georgeh0 in https://github.com/cocoindex-io/cocoindex/pull/2447 fix(surrealdb): statement errors are raised instead of swallowed; values sent as typed CBOR; string/integer/array record IDs; record writes use UPSERT ... MERGE and add TableTarget.declarefields for multi-owner records; HNSW-only vector indexes for SurrealDB 3 by @martinschaer in https://github.com/cocoindex-io/cocoindex/pull/2439 * fix(valkey): un-declaring an index or running App.drop() now also deletes its document hashes by @georgeh0 in https://github.com/cocoindex-io/cocoindex/pull/2419 * fix(valkey): CocoIndex connections are identifiable in the server's CLIENT LIST (LIB-NAME cocoindex); requires valkey-glide>=2.5.2 by @AbhayMahalle in https://github.com/cocoindex-io/cocoindex/pull/2443 fix(localfs): recursive walkdir stops at symlink cycles instead of walking forever by @Sujit-1509 in https://github.com/cocoindex-io/cocoindex/pull/2456 fix(amazons3): zero-length reads and partial reads of empty objects return empty bytes instead of sending an invalid range request by @Sujit-1509 in https://github.com/cocoindex-io/cocoindex/pull/2437 * docs(target): state that reconcile() may run more than once per commit by @georgeh0 in https://github.com/cocoindex-io/cocoindex/pull/2475 New Contributors * @YaxinCheng made their first contribution in https://github.com/cocoindex-io/cocoindex/pull/2277 * @martinschaer made their first contribution in https://github.com/cocoindex-io/cocoindex/pull/2439 * @AbhayMahalle made their first contribution in https://github.com/cocoindex-io/cocoindex/pull/2443 Full Changelog: https://github.com/cocoindex-io/cocoindex/compare/v1.0.24...v1.0.25
v1.0.242026-09-20
v1.0.24 修复了两个容易踩中的稳定性问题。一是 Linux 下 App.update() 长时间运行会永久卡死：LMDB 写事务可能被调度到另一条运行时线程上恢复，导致写锁一直被占，后续所有写入都被堵住；现在每个写事务固定在同一 OS 线程上执行。二是并发时记忆化组件被重复执行——同一 memo key 的两次 App.update() 交叠运行，后到的那次会直接复用先到的结果，不再重跑函数体；full_reprocess 下同样生效，但只在本次更新范围内复用，上一次更新留下的 memo 依旧忽略。此外补齐了容器删除语义的文档：取消声明容器或执行 App.drop() 时，CocoIndex 自建的容器连同内容一并删除，而用户自管的资源（Kafka/Igg
v1.0.232026-09-13
CocoIndex v1.0.23 主要围绕自定义目标连接器的 sink API 与若干稳定性问题。容器 sink 现在通过每次动作的子槽位提供子目标提供方（TargetActionSink.fromfnwithchildren / fromasyncfnwithchildren），直接返回 ChildTargetDef 列表的做法已废弃；Rust SDK 中对应闭包形态也迁移到了 fromasyncfnwithchilddefs，升级时需要同步改动。另外，回调相等但类型不同的 sink 不再共用同一标识，无法建立弱引用的回调（如 NamedTuple 实例）会直接抛出 TypeError。引擎侧，合并到同一 sink 调用的批次中某个组件失败时，现在只影响该组件，不再连坐全部。此前记账函数容易被误判重新执行的情况也一并修掉：重复定义旧函数、importlib.reload、函数工厂重建，以及 copy.copy / copy.deepcopy 产生的副本被丢弃，都不再清空原函数的缓存。其余两处修复：组件异常处理器现在拿到原始的 Python 异常类型，而非被包裹的 RuntimeError；tuple 和 bytes 形式的组件路径段在状态存储解码后得以保留，相应的目标状态归属不再错配或泄漏。
v1.0.222026-09-10
v1.0.22 主要围绕 LiteLLM 集成和引擎稳定性做修整。嵌入响应现在按索引与输入一一对应，修复乱序返回造成的错配；并发嵌入请求会按后端能力限流，超时改为终结性错误并新增可配置的超时参数，转写请求则统一在单个超时内重试。引擎侧，删除操作在仅降级的场景下跳过子元素存在性对账，组件活跃状态改为显式追踪，不再依赖引用计数，避免误判。CI 给构建测试设了 60 分钟上限，并实时输出 prek 钩子进度；文档站依赖同步小幅升级。
v1.0.212026-09-05
本次更新围绕数据管道的稳定性与正确性展开：引擎为 target action sink 引入基于值的身份标识，新建 target 状态时会生成全新的子 provider 代次，避免状态复用带来的隐患；neo4j 与 falkordb 在全量重处理时会正确对齐 schema 变更，Doris 则不再吞掉列 DDL 的真实错误。代码分词方面接入 tree-sitter Lua 语法，按语法结构切分更精准。CLI 修复了 Windows 应用目标路径的解析问题，事件循环在迭代器清理阶段不再被阻塞。另有三处文档补充：RecursiveSplitter 支持语言表、记忆化放置策略（组件级与函数级）、LMDB 自动扩容与逻辑指纹传播。
v1.0.202026-08-12
cocoindex v1.0.20 是一个以修复为主的小版本。核心层修了一个关于 memoization 的问题：记忆化过程中目标失效（target invalidation）信息会被丢掉，导致本该重算的内容被错误复用，现在失效状态能正确穿透缓存继续向下传递。LanceDB 集成方面，列的 schema 演进此前只在增量处理时生效，若触发全量重跑（full reprocess）则不会应用，如今两种路径行为一致，避免全量重建后表结构停留在旧版本。另外升级了依赖版本，顺带清掉了安全告警。对使用者来说，建议升级以获得更可靠的缓存失效判断和 LanceDB 全量重建的一致性。
v1.0.192026-08-04
v1.0.19 主要是修复与安全更新：依赖升级到 surrealdb 3.2.3 以上，修补了 quinn-proto 的 DoS 漏洞（CVSS 7.5）与 ammonia 的 XSS 风险；删除组件时会同步清空反向 target-owner 索引，避免残留脏数据；SQLite 列迁移失败时不再破坏 schema 一致性；memoization 的 Fingerprint 键改为确定性规范化，减少无谓的重复计算。另补了一份新示例，演示如何把会议笔记转成 SurrealDB 知识图谱，并澄清了 mount_each() 的返回语义。
v1.0.182026-07-23
cocoindex v1.0.18：cocoindex show 的目标状态检查体验明显升级。现在被追踪的 target-state 路径会以更易读的方式渲染，新增 --target-states 参数可直接列出全部目标状态，provider-only 的路径段也会持久化保存可读名称。检查稳定路径详情时改用统一解析器流式输出，速度更快。另有一项底层能力更新：code_ast 模块引入 tree-sitter 的 Dart 语法，支持按语法结构切分 Dart 代码。文档方面，安装 agent skill 改为推荐使用 npx skills CLI。
v1.0.172026-07-18
本次更新引入协同超时机制，并在 Python 侧新增 RetryWithSmallerBatch：批处理失败时自动对半拆分重试，降低整体失败率。修复方面，LMDB 读事务现会提交以保证跨进程句柄有效，Postgres 绑定中的 NUL 字符（U+0000）也会递归清除，避免数组与复合类型写入出错。新增「Slides to Speech」示例，支持本地 CPU 语音合成与逐页增量处理、带记忆化，全程无需 TTS API key。安全上将 pyo3 从 0.27 升级至 0.29，并同步更新依赖以修复已知漏洞；另补充了 RateLimiter 限流文档。