RAG Pipeline Builder

Part of the worlds-biggest-software-project initiative.

A framework-agnostic visual builder for retrieval-augmented generation pipelines, with built-in evaluation and clean code export.

RAG Pipeline Builder is an open-source platform for designing, deploying, and evaluating retrieval-augmented generation pipelines. It targets AI engineers, platform teams, and data teams who need a faster path from raw documents to a production-quality RAG service without locking themselves into a single orchestration framework or managed cloud.

Why RAG Pipeline Builder?

Visual builders today are framework-locked. deepset Studio is the only serious visual RAG builder with production deployment, but it is tied exclusively to Haystack. No equivalent exists that targets LlamaIndex, LangChain, and Haystack components on equal footing.
Chunking is hand-tuned and opaque. Chunking strategy is the single biggest quality lever in RAG, yet no tool benchmarks strategies against a sample query set and recommends an optimal approach.
Evaluation lives in a separate tool. Most builders require RAGAS or LangSmith wired in externally; the build-evaluate-improve loop is fragmented across products, slowing iteration.
Documents are treated as blobs. None of the surveyed tools understands PDF table structure, spreadsheet schemas, or code semantics, leaving retrieval precision on the table.
Enterprise access control is bolted on. Multi-tenancy with namespace isolation and row-level access control at the retrieval layer is implemented ad-hoc by every team rather than offered as a first-class configuration option.

Key Features

Pipeline Authoring

Drag-and-drop visual canvas with bidirectional code export to Python and/or TypeScript
Hybrid retrieval (dense embeddings plus BM25) with reciprocal rank fusion
Configurable chunking (fixed-size, sentence, recursive) with overlap control
Document ingestion pipeline supporting PDF, Word, HTML, Markdown, and plain text
Pre-built templates for standard RAG, agentic RAG, and hybrid retrieval patterns

Evaluation In The Loop

Built-in RAGAS-compatible metrics: faithfulness, answer relevancy, context precision, context recall
Batch evaluation runner against a golden query set, runnable from the same UI
Automated chunking strategy benchmarker that scores multiple strategies on sample queries
Per-query trace showing query, retrieved chunks, generated answer, latency, and token counts

Agentic And Multimodal Capabilities

Agentic RAG mode with query grading, document relevance scoring, and iterative query reformulation
Multimodal ingestion: table and image extraction from PDFs with vision model captioning
Optional Graph RAG overlay for multi-hop retrieval over a knowledge graph
Knowledge gap detection to surface query topics not covered by the indexed corpus

Deployment And Integration

REST API for ingest, query, and evaluate operations with auto-generated OpenAPI specs
MCP server export so any pipeline can be deployed as a tool server for agent ecosystems
Multi-tenant workspaces with namespace isolation and configurable RBAC at the retrieval layer
Cost and latency profiler with per-query breakdown across embedding, reranker, and LLM stages

AI-Native Advantage

AI is used to close the loop that today requires human glue between separate tools. A chunking advisor benchmarks strategies against real queries instead of relying on intuition. Query reformulation (HyDE, step-back prompting) runs automatically before retrieval. RAGAS evaluation is triggered on document or pipeline changes to surface regressions instantly, and embedding-coverage analysis identifies gaps between what users ask and what the corpus contains.

Tech Stack & Deployment

The project targets self-hosted deployment via Docker or Kubernetes alongside an optional managed cloud. It aligns with established standards: OpenAPI for service exposure, LangChain Expression Language for pipeline interoperability, MTEB for embedding model selection, and RAGAS for evaluation. Components from LlamaIndex, LangChain, and Haystack are intended to be usable within the same pipeline graph rather than forcing a single-framework choice.

Market Context

The global RAG market was valued at approximately USD 2.33 billion in 2025 and is projected to reach USD 3.33 billion in 2026, scaling toward USD 81.51 billion by 2035 at a CAGR of 42.7% (NextMSC, 2026). Pricing across incumbents spans free self-hosted open source through managed vector databases at USD 70–300/mo and visual builder SaaS at USD 100–500/mo, with enterprise platforms reaching USD 100K–200K+ annually (AlphaCorp, 2026). Primary buyers are AI engineers building knowledge-base assistants, platform teams standardising the internal RAG stack, and enterprise architects replacing legacy search with semantic retrieval.

Project Status

This project is in the research and specification phase.
Contributions, feedback, and domain expertise are welcome.

Contributing

We welcome contributions from developers, domain experts, and potential users. See CONTRIBUTING.md for guidelines.

Important: All contributions must be your own original work or clearly attributed open-source material with a compatible licence. Copyright infringement and licence violations will not be tolerated and will result in immediate removal of the offending contribution. If you are unsure whether a piece of code, text, or other material is safe to contribute, open an issue and ask before submitting.

Licence

Licence to be determined. See discussion for context.