Verba for RAG
Weaviate's open-source RAG app — a working reference, and what it is and isn't built for.
Verba was Weaviate’s open-source modular RAG application — the Golden RAGtriever — built to demo end-to-end retrieval-augmented generation on top of a Weaviate backend. As of 28 July 2026, the weaviate/Verba GitHub repository is archived (8 June 2026) and Weaviate’s official blog marks the project deprecated, pointing new interactive demos to playground.weaviate.io and Elysia instead.
What was Verba, and what part of a RAG system did it own?
Verba was an end-to-end RAG interface plus modular pipeline: ingest documents through a web UI, chunk and embed them into Weaviate, retrieve with hybrid search, then generate answers with source chunks highlighted (Weaviate engineering blog; archived GitHub README). It owned the application and modular RAG stages, not a separate vector engine — retrieval quality still lived in Weaviate’s indexing, hybrid fusion and chunking choices. The live GitHub capture on 2026-07-28 shows about 7,700 stars. Deep vector-database mechanics belong on Weaviate for RAG; this page covers the reference app only.
How did Verba structure a RAG pipeline?
Verba broke RAG into five swappable managers documented in Weaviate’s architecture post: Reader (ingest paths and URLs into Verba documents) → Chunker (split text, default overlap chunker) → Embedding (vectorise into Weaviate) → Retrieve (hybrid search with features such as autocut) → Generation (LLM answer from retrieved chunks). That modularity was the reference value: teams could swap readers, chunkers or generators without rewriting the whole stack. Hybrid fusion detail is at hybrid search; chunking trade-offs at chunking for RAG.
Which models and data sources did Verba support?
The archived GitHub feature tables list broad provider support for both generation and embeddings — including Ollama, Hugging Face, OpenAI, Cohere, Anthropic, Groq and others — plus ingestion paths for PDF, DOCX, CSV, GitHub/GitLab, Unstructured, Firecrawl and AssemblyAI transcription. A Weaviate-native embedding path was included alongside third-party embedders. The README warns that mixing Ollama embedding models with different vector dimensions causes errors, which is a real operational footgun when swapping models on the same index.
What were Verba’s capabilities and limits, side by side?
Verba was designed as a working reference, not a full production platform. The table below uses pre-discontinuation sources and keeps the limits that matter for anyone evaluating it today.
| Capability | What you got | The limit that rode along |
|---|---|---|
| Modular five-manager pipeline | Swap readers, chunkers, retrievers and generators independently | Storage model was Weaviate-centric — not a general vector-database adapter framework |
| Hybrid retrieval plus source UI | Hybrid search with chunk highlights and document drill-down | Reranking remained planned, not shipped, in the README feature table at discontinuation |
| Low-friction ingestion UI | Upload documents without writing ingestion scripts | Not equivalent to enterprise knowledge platforms with SLAs and multi-tenant governance |
| Local demo path | Ollama and local Weaviate for private RAG experiments | Embedding-model changes could break indexes when vector dimensions differ |
| Open-source reference | About 7,700 GitHub stars at the live capture | Archived 8 June 2026 — no bug fixes, security patches or feature work since |
| Explicit scope boundaries | README marked agentic RAG and graph RAG as out of scope | Not built for agent loops or graph-native retrieval patterns |
| Roadmap items | Advanced querying, reranking and RAG evaluation listed as planned | Planned features were still open when the project was discontinued |
The topical-map line is deliberate: a working reference, and what it is and isn’t built for. Verba was excellent for teaching modular RAG on Weaviate. It was never positioned with the same maintenance rigor as Weaviate production products — the archived README said so explicitly before the archive notice.
What is Verba’s status as of July 2026?
Verba is discontinued. The GitHub repository was archived on 8 June 2026 with an owner notice that no further updates, bug fixes, security patches or features will arrive; pull requests and issues are no longer reviewed. Weaviate’s blog now carries a deprecation banner directing new demos to playground.weaviate.io, including Elysia as a next-generation agentic RAG assistant. Existing deployments may continue to run as-is, but new production bets should not assume maintenance.
When should you have chosen Verba — and what should you use instead now?
Verba fit quick Weaviate-native RAG demos, teaching modular pipelines, and local Ollama experiments where a UI mattered more than enterprise controls. It did not fit agentic RAG, graph RAG, or teams needing shipped reranking and evaluation interfaces — those were out of scope or unfinished in the README.
For new work in July 2026, treat Verba as historical reference code. Maintained alternatives depend on the job: build directly on Weaviate and its docs for retrieval-first apps; use Dify or RAGFlow for batteries-included platforms; use LangChain or LlamaIndex for library-first control; use Weaviate’s playground or Elysia for vendor-hosted demos. The scored framework branch verdict remains at choosing a RAG framework.
What was Verba?
Verba was Weaviate's open-source modular RAG application — the Golden RAGtriever — built to demo end-to-end retrieval-augmented generation on top of a Weaviate backend. It owned the application and modular RAG stages (Reader → Chunker → Embedding → Retrieve → Generation), not a separate vector engine.
Is Verba still maintained?
No. The weaviate/Verba GitHub repository was archived on 8 June 2026 with an owner notice that no further updates, bug fixes, security patches or features will arrive. Weaviate's blog marks the project deprecated. Existing deployments may continue to run as-is, but new production bets should not assume maintenance.
How did Verba use Weaviate?
Verba chunked and embedded documents into Weaviate, then retrieved with hybrid search (including features such as autocut) before generating answers with source chunks highlighted. Retrieval quality still lived in Weaviate's indexing, hybrid fusion and chunking choices — Verba was the reference UI and modular pipeline around that backend.
Could Verba run locally with Ollama?
Yes. The archived feature tables list Ollama alongside local Weaviate for private RAG experiments. The README warns that mixing Ollama embedding models with different vector dimensions causes errors — a real operational footgun when swapping models on the same index.
What replaces Verba?
For vendor-hosted demos, Weaviate points to playground.weaviate.io and Elysia. For new builds, work directly on Weaviate for retrieval-first apps; use Dify or RAGFlow for batteries-included platforms; use LangChain or LlamaIndex for library-first control. The scored framework verdict remains at /decisions/framework/.