Weaviate’s kitchen-sink RAG app ships with a warning label

What it does Verba is a web-based personal assistant for Retrieval-Augmented Generation. It ingests documents—PDFs, spreadsheets, code repos, audio transcriptions, and web pages—chunks them using techniques from spaCy token splitting to semantic grouping, and answers questions via a chat interface. You can run it entirely offline through Ollama and HuggingFace, or pipe requests to OpenAI, Anthropic, Cohere, Groq, and others.

The interesting bit The project treats RAG as a configurable pipeline rather than a fixed stack. It exposes chunking strategies for specific formats like Markdown, JSON, and code, and includes a 3D vector viewer to peek into how your data is embedded. That granularity is unusual for a tool that also ships as a one-click pip package.

Key highlights

• Broad model support: local models via Ollama and HuggingFace, plus cloud providers including Anthropic, OpenAI, Cohere, Groq, Novita AI, and Upstage.
• Pluggable data ingestion: PDF, DOCX, CSV/XLSX, GitHub/GitLab repos, URLs via Firecrawl, and audio via AssemblyAI.
• Eight chunking modes, including semantic similarity grouping and recursive rules, powered by spaCy.
• Hybrid search combining semantic and keyword retrieval, with autocomplete and metadata filtering.
• LangChain integration for RAG pipelines; Docker and customizable frontend included.

Caveats

• Weaviate explicitly warns that Verba is a community project and “may not be maintained with the same rigor as production software.”
• Several advanced features—reranking, RAG evaluation, Haystack and LlamaIndex support—are marked as planned but not yet implemented.
• Agentic RAG, Graph RAG, and multi-user collaboration are currently out of scope.

Verdict Good for developers who want a ready-made RAG UI to prototype document chat without wiring up vector databases and chunking logic manually. Skip it if you need production-grade reliability, multi-user workflows, or agentic pipelines.