v1.0.02026-11-06
After three release candidates, we are pleased to announce the official release of RAGFlow 1.0, now ready for production. Built on a complete rewrite in Go, this release significantly reduces deployment and runtime overhead and introduces an all-new Agentic Chat experience. We encourage all users to upgrade for improved stability, performance, and question answering.
Highlights
- Rewritten in Go for lighter deployment: Docker images are approximately 10× smaller, and startup is approximately 10× faster.
- Approximately 100× greater concurrency, with substantial improvements in query latency, memory usage, and document parsing speed.
- All-new Agentic Chat: Enables multi-step reasoning and answers to complex questions through iterative retrieval, tool calls, and evidence sufficiency assessment.
Chat and Agents
- Added Agentic Chat mode, replacing the previous Low, Medium, High, and Ultra thinking levels.
- Introduced the Jev decision model into the agent tool-calling loop to determine whether the available evidence is sufficient to answer a question.
- Added support for the Jev decision model in the Categorize agent component.
- Added storage and display of prompts, token usage, and execution time in chat history.
Document Ingestion and Retrieval
- Unified the 12 built-in document parsers with the pipeline parsers.
- Added support for multiple ingestion pipeline DSL versions.
Architecture and Deployment
- Migrated the message queue from Redis to NATS.
- Adopted Kvrocks for cache and checkpoint storage.
- Added Docker images for the arm64 architecture.
- Reimplemented the RAGFlow CLI in Go.
Stability
- Fixed multiple critical bugs reported by the community and issues identified during the release candidate phase, further improving runtime stability.
Benchmarks
- Achieved 91% accuracy on the FRAMES multi-hop question-answering benchmark using RAGFlow Agentic Chat with MiniMax M3.
Upgrading
- From v1.0.0-rc1, rc2, or rc3: Existing data is fully compatible. You can upgrade directly.
- From v0.27.2: You can upgrade directly to v1.0.0. Data migration runs automatically when the RAGFlow Docker container starts.
- From versions earlier than v0.27.2: Follow the migration guide to complete the required intermediate upgrades before upgrading to v1.0.0.
- Back up all data before upgrading. Data migration from v0.27.2 to v1.0.0 is irreversible. If migration fails, restore the complete pre-upgrade backup before trying again.
Compatibility and Limitations
- Deprecated API aliases are no longer supported.
- The Python SDK is unaffected.
- DeepDoc currently supports CPU environments only.
Relative references
- Backup and upgrade guide: https://ragflow.io/docs/migration_overview
- Removed APIs: https://ragflow.io/docs/httpapireference#removed-api-aliases
What's Changed
* fix(parser): keep recovered PDF TOC leader lines out of the body guard by @jay77721 in https://github.com/infiniflow/ragflow/pull/21768
* CI: Fix log tee failure by @myf-bee in https://github.com/infiniflow/ragflow/pull/21791
* Update release notes of 1.0.0.rc3 by @yingfeng in https://github.com/infiniflow/ragflow/pull/21793
* feat: add Wally as a model provider by @adityabatra072 in https://github.com/infiniflow/ragflow/pull/21719
* add audit hooks to file APIs by @JinHai-CN in https://github.com/infiniflow/ragflow/pull/21807
* add log hooks by @JinHai-CN in https://github.com/infiniflow/ragflow/pull/21822
* i18n: translate pipeline version and parser settings strings in tr-TR by @bakiburakogun in https://github.com/infiniflow/ragflow/pull/21819
* fix(models): honour context cancellation in driver HTTP requests by @ccclucky in https://github.com/infiniflow/ragflow/pull/21643
* feat(vastbase): support Vastbase G100 as the metadata database in Go by @FengYikai710 in https://github.com/infiniflow/ragflow/pull/20913
* fix: README.md by @eridan-ltu in https://github.com/infiniflow/ragflow/pull/21828
* fix(chunker): prevent PDF child chunks from overwriting each other by @jay77721 in https://github.com/infiniflow/ragflow/pull/21773
* test(sdk): remove explicit Python SDK test skips by @6ba3i in https://github.com/infiniflow/ragflow/pull/21783
* fix(syncer): retain Dropbox checkpoints on download failure by @Zhi-Xiang-Guo in https://github.com/infiniflow/ragflow/pull/21809
* fix(api): keep the datasets API error response body shape by @simpleqt in https://github.com/infiniflow/ragflow/pull/21831
* perf(web): virtualize chat message lists by @dcc123456 in https://github.com/infiniflow/ragflow/pull/21790
* CI: add check for ASCII, disallow none ASCII code check-in by @wangq8 in https://github.com/infiniflow/ragflow/pull/21784
* fix(anthropic): support struct json schema in tool parameters by @Sanderhoff-alt in https://github.com/infiniflow/ragflow/pull/21638
* feat(parser): hand off Markdown and HTML remote images as media refs by @jay77721 in https://github.com/infiniflow/ragflow/pull/21764
* fix(paddleocr): handle submit request creation errors by @QAQadws in https://github.com/infiniflow/ragflow/pull/21839
* fix(deepdoc): add cgo constraints to native configuration tests by @QAQadws in https://github.com/infiniflow/ragflow/pull/21837
* fix: SoMark web page is missing by @fenglingxiansheng in https://github.com/infiniflow/ragflow/pull/21710
fix(agent): clamp dataset navigation maxdocs to the default budget by @flyingwhit in https://github.com/infiniflow/ragflow/pull/21755
* feat(admin): add user search, status filtering, sorting, and CLI pagination by @QAQadws in https://github.com/infiniflow/ragflow/pull/21729
* feat(chat): add AnySearch web search provider by @GunaPalanivel in https://github.com/infiniflow/ragflow/pull/20633
* refa: consolidate model access and tenant defaults by @buua436 in https://github.com/infiniflow/ragflow/pull/21740
* fix(ingestion): stop MinerU JSON path from inlining zip figure bytes by @Souravrajvi0 in https://github.com/infiniflow/ragflow/pull/21803
fix(tool): cap tool HTTP response bodies at 4 MiB in arxiv, querit, duckduckgo and yahoofinance by @flyingwhit in https://github.com/infiniflow/ragflow/pull/21769
* fix(models): publish the provider catalog through an atomic pointer and drop the dao-side cache so concurrent readers and publishers no longer race by @flyingwhit in https://github.com/infiniflow/ragflow/pull/21749
fix(agent): resolve DeepL and QWeather credentials from node config or DEEPLAPIKEY/QWEATHERAPI_KEY instead of model-visible params by @flyingwhit in https://github.com/infiniflow/ragflow/pull/21735
* fix(dify): keep retrieval unscoped when metadata conditions are empty by @MoonPointer-Byte in https://github.com/infiniflow/ragflow/pull/21766
* test(retrieval) : add unit tests for retrieval and reranker by @zzy1017-crypto in https://github.com/infiniflow/ragflow/pull/21611
* fix(agent): prevent checkpoint crashes and propagate component errors by @6ba3i in https://github.com/infiniflow/ragflow/pull/21854
* fix(agent): detect webhook Begin in saved canvases by @6ba3i in https://github.com/infiniflow/ragflow/pull/21849
* fix(file): record commit deletions without a missing status column by @ccclucky in https://github.com/infiniflow/ragflow/pull/21796
* fix(tokenizer): replace the default fields instead of appending by @ccclucky in https://github.com/infiniflow/ragflow/pull/21799
* test(mediacache): cover real Kvrocks lifecycle and index boundary by @jay77721 in https://github.com/infiniflow/ragflow/pull/21851
* fix(user): use string for user id in DAO and handler lookups by @xugangqiang in https://github.com/infiniflow/ragflow/pull/21792
* fix(build): adjust onnxruntime static link flags on Darwin arm64 by @Sanderhoff-alt in https://github.com/infiniflow/ragflow/pull/21162
* release dockerhub image remove unnecessary tag by @myf-bee in https://github.com/infiniflow/ragflow/pull/21860
* fix(chunk): number paragraph ch
v1.0.0-rc32026-10-30
1.0.0-rc3 是 v1.0.0 正式版前的最后一个候选版,修复了 rc2 暴露的一批问题,并补齐了几项功能。智能体工具循环中新增证据充分性判定的决策模型,同时引入 Categorize 组件;数据摄入管道的 DSL 支持多版本并存;聊天记录现在会保存提示词、token 消耗与执行耗时。向上兼容 rc2 的数据结构,可直接升级。
解析侧改动较多:标题族切分保留结构化表格与图片,图片摄入可提取 SVG 文本,PDF 表格统一走 RenderTableHTML 渲染,DocItem 位置改用类型化的 Position API。S3 兼容存储开始遵循 signatureversion 与 addressingstyle 配置,Go 客户端补齐 Mistral OCR 接口。智能体方面修复了模板多出检索节点、引用内容显示异常、webhook 头部字段丢失与增量轮询中断等问题,agentic RAG 也接入了 token 用量统计。另有权限字段迁移、元数据字符串数组过滤、删除数据集元数据等十余项修复。
v0.27.32026-10-23
这版主要围绕 Go 版重写和 MCP 检索链路做修补。agentic RAG 相关 PR 已批量移植到 Go,并补上集成测试;MCP 侧修复了 Info() 不翻译 inputSchema.properties、以及没传 rerankcandidatescount 导致检索翻页超过 64 条就失效的问题。知识库比对改用解析后的 embedding 模型,避免同一模型因 embd_id 不同被误判为不同库。
其余修复分散在各处:取消数据集任务需按权限校验,URL 防护解开 IPv6 过渡地址以堵住 SSRF,S3 瞬时故障加入重试,嵌套编译器池的 embedding 死锁被消除,ONNX Runtime 统一到 1.29.0。界面与前端也有一批小修——DOCX 预览不再陷入无限更新循环,wiki 修改标记误报、TreeSelect 与下拉框的显示问题、繁体中文分隔符翻译均已处理;pipeline 里的解析算子文件现在可以删除了。
v1.0.0-rc22026-10-16
RAGFlow 1.0.0-rc2 是 v1.0.0 的第二个候选版,修复了 rc1 的一批问题,并补齐若干能力。最大的变化是聊天引入 Agentic 模式,取代原先的 Low/Medium/High/Ultra 四档思考级别;rc1 的数据完全兼容,可直接升级。新增 Team/Me 分享功能,接入模型服务商 DemonRoute 与 Opper、联网搜索工具 Firecrawl 与 Search1API,以及 MinerU 4 模型;捷克语、斯洛伐克语的检索支持也已加入。文档解析方面,12 个内置解析器与 Pipeline 解析器完成统一,表格统一走分段 HTML 输出,原生 PDF/Office 抽取收拢到 deepdoc 门面。稳定性上,Elasticsearch 现已支持任意维度的向量,修复了分词符配置无法删除最后一项、解析配置因别名共享被串改等 bug,模型调用统一经由 wrapper 路由。此外提供 arm64 架构 Docker 镜像。
v1.0.0-rc12026-09-29
RAGFlow v1.0.0-rc1 是 1.0.0 的预览版,核心变化是用 Go 对整个项目做了重写:资源占用大幅下降,原先偏重的部署方式改为轻量架构。消息队列从 Redis 迁到 NATS,缓存与检查点统一交给 Kvrocks 承担。
升级前务必先备份数据,且确认当前版本为 v0.27.2——容器启动时会自动完成数据迁移,该过程不可逆,升级失败只能从备份恢复并提 issue。
兼容性方面:RAGFlow CLI 已用 Go 重写,部分已废弃的 API 不再支持;Python SDK 不受影响。已知限制包括 DeepDoc 仅支持 CPU 环境、本地沙箱被移除、Team/Me 权限模型尚未支持。
v0.27.22026-09-10
RAGFlow 本次更新主打检索与解析。Agentic RAG 检索框架重构后,推理速度与基准测试成绩明显提升;知识编译新增运行时设置并加入速率限制,大批量入库更可控。知识图谱页面现在显示节点总数与当前数量,输入实体名回车即可定位并高亮。接入侧新增 Sitemap 数据源,可按 sitemap.xml 抓取网页,WebDAV 连接器支持自定义 CA 证书。解析新增 MonkeyOCRv2 与自部署 PaddleOCR-VL 支持,EPUB 文档可预览,Excel 引用来源也能在预览中定位。Agent 新增 Sofya 搜索工具,模型供应商接入 Hubris 与 llmman。另修复聊天中上传图片被视觉模型忽略、CSV/TXT 非 UTF-8 编码解析失败等问题,并新增 RERANKTOKENLIMIT_MODE 配置。
v0.27.12026-08-28
RAGFlow 本次更新先看新能力:数据源接入 Azure DevOps 连接器,聊天与 Agent 新增 You.com、Serply 两家网络搜索;模型侧加入 Synthorai 提供商,补齐 DeepSeek 缺失模型,AWS Bedrock 支持 API key 认证。检索 API 开放 rerankcandidatescount、knn topk、numcandidates 等参数,元数据过滤下推至索引层,检索更快;聊天设置保存出错时会自动滚动到报错位置,搜索页则会校验知识库是否已被删除。修复面较广:PDF 解析失败、MinerU 渲染与代码块解析、QA Excel 丢行(答案为 0)、编码检测误弃结果、OCR 文本丢失、BigQuery SQL 拼接错误、WebDAV 清理导致数据丢失、跨知识库文档 ID 冲突,以及聊天自动滚动不跟随流式输出等问题一并解决。
v0.27.02026-08-19
v0.27.0 带来两项重头更新:全新的知识编译功能,支持文档级与数据集级编译,产出 Wiki、Graph、Tree、Page Index、思维导图、时间线等多种形态;原先的 GraphRAG 与 RAPTOR 已下线,界面中不再提供,由 Graph、Tree 两个新模块接替,已生成的旧内容仍可正常检索。另一项是全新的 Agentic RAG,回答时提供 Low、Medium、High、Ultra 四档思考深度可选。
模型提供方系统整体重构,配置与管理更顺手;新增 Qwen 3.8 系列、Kimi K3、OpenRouter Embedding、AWS Bedrock Reranker、Mistral OCR 等一批模型与接口支持。底层加入 GaussDB、SereneDB 适配,Infinity 升至 0.7.3;另新增 Querit 联网搜索工具,并优化了 Markdown 排版与搜索、聊天的筛选功能。
修复面较广,重点包括:解析大文件卡死、上传拖垮整个服务、删除数据集阻塞其他请求等稳定性问题;以及多模型对话忽略所选模型、会话切换残留旧消息、SSE 流中断、中文文件名 PDF 预览失败、CSV 表格解析丢字段等具体缺陷。
v0.26.42026-07-07
RAGFlow 于 7 月 7 日发布新版本。分词能力升级:新增基于 Snowball 的多语言词干提取器,覆盖 16 种语言,分词全链路接入 language 参数,前端也补上了荷兰语。同时加入了知识编译工作流。
修复方面较为密集:解析含 @ 的 LM-Studio 模型名不再崩溃;MCP 服务器的 list_chats 不再因接口返回分页字典而报错;Docling 解析器此前会悄悄丢弃数学公式,现在能正常抽取。元数据的行内编辑能持久化到后端,批量挂载文件到知识库不再覆盖已有链接,中文过滤器失效问题也已修复。另外还处理了标签权重未强制大于零、Agent 提示词下拉菜单错位、Laws 分块器把点号交叉引用误判为标题,以及 MCP 服务器重名校验、空选区导出等一系列表单校验漏洞。
v0.26.32026-07-02
RAGFlow 于 2026 年 7 月 2 日发布新版。数据源新增 Google BigQuery 连接器,支持文档接入与增量同步;MCP 服务端加入 ragflowlistdatasets 和 ragflowlistchats 两个工具。文档解析集成了版面感知的 SoMark OCR 解析器,能更准确抽取表格、图形等复杂元素;API 同时开放 Ingest documents 接口,可按自定义管线处理文档。
使用体验上有两处改进:批量上传改为部分成功模式,单个文件失败不再导致整批丢弃;全局导航栏会根据屏幕宽度在桌面横栏与移动端抽屉之间切换,重叠问题得以解决,手机端明显更好用。
修复的问题包括:配置 Ollama 后报 'Failed to access model(name) using this api key'、删除消息时因索引错位误删上一轮对话引用、DOCX 非编号标题样式解析失败,以及按 embd_id 取不到嵌入模型等。