RAGFlow 1.0预览：从检索引擎转向Agent原生数据库
RAGFlow 1.0 preview: moving from RAG engine to agent-native database
RAGFlow提出将检索视为Harness问题而非索引问题的新范式，用类型化工件替代传统编译优化，这种工程取舍对Agent开发极具参考价值，建议关注其架构设计思路。
Context as Infrastructure: Knowledge Compilation for Retrieval in Agentic Systems with Harnesses
作为基础设施的上下文：在带有 Harnesses 的智能体系统中为检索进行知识编译
📄 https://arxiv.org/abs/2610.41172
Sharing the paper behind RAGFlow 1.0.
分享 RAGFlow 1.0 背后的论文。
RAG’s center of gravity has moved. Retrieval is now a loop: the agent queries, reads, reformulates, and decides when it has enough. Once that loop exists, better chunking/embeddings/Recall@k stop being the main bottleneck. The agent can find things. The problem is the cost of reading.
RAG 的重心已经转移。检索现在是一个循环：智能体进行查询、阅读、重新表述，并判断何时已获得足够信息。一旦形成该循环，更好的分块/嵌入/Recall@k 就不再是主要瓶颈。智能体能够找到所需内容，问题在于阅读的代价。
Two conclusions from the paper:
论文得出的两个结论：
- Retrieval is a harness problem before it’s an index problem. The index decides what can be found; the harness decides whether finding it becomes an answer.
- Compilation is not a better retriever, and doesn’t need to be. We tested summary trees, outlines, wikis, and graphs under a competent harness. None gave a detectable accuracy gain. Structure isn’t extra evidence. What compilation can do is supply better context: smaller, provenance-bound, and sufficient.
- 在成为索引问题之前，检索首先是一个 Harness 问题。索引决定了什么可以被找到；Harness 决定了找到它是否能转化为答案。
- 编译并不是为了获得更好的检索器，也不需要如此。我们在具备能力的 Harness 下测试了摘要树、大纲、维基和图结构。均未带来可检测的准确率提升。结构并非额外的证据。编译所能做的是提供更好的上下文：更精简、具有来源绑定且信息充足。
So we stopped treating compilation as retrieval and started treating it as a supply layer. At ingestion time, docs are turned into typed artifacts—claims with verbatim evidence, relations, timelines, etc. The harness then decides which artifacts enter the context window and when to stop. Context becomes infrastructure when it’s compiled once, typed, versioned, and governed.
因此，我们不再将编译视为检索手段，而是将其视为供应层。在摄入阶段，文档被转换为类型化的工件——如带有逐字证据的主张、关系、时间线等。然后由 Harness 决定哪些工件进入上下文窗口，以及在何时停止。当上下文经过一次性编译、类型化、版本化和治理时，它就成为了基础设施。
The tradeoff is explicit: we save on reading cost, but pay for it in extra retrieval rounds. For claims, graph, and wiki, we published an empty results ledger—metrics and decision rules are listed, but every row is marked “no measured effect” because none has met its pre-registered target yet. An artifact the harness never uses is just cost, not context. I’d rather say that than fill the column with claims from demo collections.
这种权衡是显式的：我们节省了阅读成本，但为此付出了额外检索轮次的代价。对于主张、图和维基，我们发布了一份空结果账本——列出了指标和决策规则，但每一行都标记为“无测量效果”，因为目前没有任何一项达到其预注册的目标。Harness 从未使用的工件仅仅是成本，而非上下文。我宁愿这样说，也不愿用演示集合中的声明来填充这一列。
Happy to discuss.
欢迎讨论。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力