PipesHub
An open-source project focused on enterprise knowledge retrieval and RAG pipeline construction, providing scalable solutions for retrieval-augmented generation systems.
Timeline (last 90 days)
The experiment found that a small reranker actually reduced the accuracy of the best pipeline by 9 percentage points, and switching to a larger reranker yielded almost no improvement.
All pipeline variants used the same model, the same embedding vectors, and the same document corpus; the only variables were the retrieval and orchestration strategies.
In the Google FRAMES benchmark, a simple agent loop (capable of reading retrieved results and searching again) achieved an accuracy of 92.7%.
This test covered all 824 multi-hop questions in FRAMES and built 18 pipeline variants to break down the contribution of each component.
The benchmark code and runbook have been made publicly available in the PipesHub repository, and the conclusions can be independently reproduced.
The team performed a secondary verification on every correct answer, comparing whether the content the system actually read was consistent with the answer.
PipesHub is an open-source Context Layer licensed under Apache 2.0.
A PipesHub developer shared on Reddit their experience building the project over an extended period, noting that enterprise data is often scattered across S3, Google Drive, Slack, Jira, Confluence, SharePoint, email, and various databases.
PipesHub supports connecting to the user's choice of LLM and Embedding models, including locally deployed models.
PipesHub is an open-source context layer under the Apache 2.0 license, sitting between enterprise data and upper-layer AI applications, responsible for organizing scattered company data into unified context consumable by search, conversation, Agents, MCP clients, or self-developed applications.
10 more timeline events
All Facts (20)
In the Google FRAMES benchmark, a simple agent loop (capable of reading retrieved results and searching again) achieved an accuracy of 92.7%.
50%UnverifiedThe experiment found that a small reranker actually reduced the accuracy of the best pipeline by 9 percentage points, and switching to a larger reranker yielded almost no improvement.
50%UnverifiedAll pipeline variants used the same model, the same embedding vectors, and the same document corpus; the only variables were the retrieval and orchestration strategies.
50%UnverifiedThe team performed a secondary verification on every correct answer, comparing whether the content the system actually read was consistent with the answer.
50%UnverifiedThis test covered all 824 multi-hop questions in FRAMES and built 18 pipeline variants to break down the contribution of each component.
50%UnverifiedThe benchmark code and runbook have been made publicly available in the PipesHub repository, and the conclusions can be independently reproduced.
50%UnverifiedPipesHub is an open-source Context Layer licensed under Apache 2.0.
50%UnverifiedPipesHub supports connecting to the user's choice of LLM and Embedding models, including locally deployed models.
50%UnverifiedPipesHub is an open-source context layer under the Apache 2.0 license, sitting between enterprise data and upper-layer AI applications, responsible for organizing scattered company data into unified context consumable by search, conversation, Agents, MCP clients, or self-developed applications.
50%UnverifiedPipesHub supports self-hosted deployment and can be fully deployed on an enterprise's own infrastructure, with data never leaving the enterprise's internal network.
50%UnverifiedPipesHub provides permission-aware retrieval, preserving the permission system of the data source itself to ensure retrieval results comply with user permissions.
50%UnverifiedPipesHub combines knowledge graphs with semantic retrieval, not relying solely on vector similarity matching but also introducing graph structures to enhance contextual understanding.
50%UnverifiedCore engineering challenges in enterprise-grade RAG include permission preservation, document change detection, cross-source deduplication, and precise citation traceability—issues that are almost imperceptible at the prototype stage.
50%UnverifiedPipesHub's core features include self-hosted deployment, permission-aware retrieval, precise citation traceability, a combination of knowledge graphs and semantic retrieval, flexible model integration, multi-language SDK support (Python, TypeScript, Go), and MCP protocol invocation.
50%UnverifiedPipesHub is an open-source context layer licensed under Apache 2.0, positioned as a middleware layer between enterprise data and upper-level AI applications.
50%UnverifiedPipesHub source code is hosted on GitHub (github.com/pipeshub-ai/pipeshub-ai).
50%UnverifiedPipesHub supports a variety of pluggable underlying technology choices, including Neo4j and ArangoDB as graph databases, Qdrant, OpenSearch, and Redis as vector databases, and Kafka and Redis Streams as message queues.
50%UnverifiedPipesHub's core features include self-hosted deployment, permission-aware retrieval, precise citation traceability, a combination of knowledge graphs and semantic retrieval, and flexible model integration.
50%UnverifiedPipesHub provides Python, TypeScript, and Go SDKs, as well as MCP protocol call support.
50%UnverifiedA PipesHub developer shared on Reddit their experience building the project over an extended period, noting that enterprise data is often scattered across S3, Google Drive, Slack, Jira, Confluence, SharePoint, email, and various databases.
50%