Verba — The Golden RAGtriever by Weaviate

Introduction

Verba is the polished, opinionated RAG application open-sourced by Weaviate. It bundles document ingestion, chunking, embedding, vector search, and chat into a single application — making "talk to your documents" a 2-minute setup instead of a multi-week project.

With over 7,500 GitHub stars, Verba is a great first-RAG-app for teams who want to evaluate RAG patterns, demo to stakeholders, or self-host a private knowledge assistant without writing code.

What Verba Does

Upload documents (PDF, DOCX, Markdown, plain text, web URLs, GitHub repos). Verba chunks them, embeds with your chosen embedder (OpenAI, Cohere, Ollama, Sentence Transformers), stores vectors in Weaviate, and serves a chat UI that retrieves relevant chunks and prompts the configured LLM (OpenAI, Anthropic, Ollama, OpenRouter, etc.).

Architecture Overview

[Verba Frontend (Next.js)]
      |
[Verba Backend (FastAPI)]
   pluggable Reader / Chunker / Embedder / Generator
      |
[Weaviate Vector DB]
   semantic search over chunks
      |
[Embeddings + Generation Providers]
   OpenAI, Cohere, Ollama, Anthropic, Mistral,
   Sentence Transformers, OpenRouter, ...
      |
UI flow:
  Upload docs -> chunked -> embedded -> stored
  Ask question -> embed -> retrieve top-k chunks -> LLM answer with citations

Self-Hosting & Configuration

# .env (configure providers)
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
OLLAMA_URL=http://host.docker.internal:11434
OLLAMA_MODEL=llama3.1

WEAVIATE_URL_VERBA=http://weaviate:8080   # external Weaviate
# Or leave empty to use Weaviate Embedded (no infra)

# Common workflows
verba start --port 8000             # local UI
verba reset                         # wipe stored data
# Docker Compose with Weaviate + Verba + Ollama
#   profiles: ollama, openai, default — pick at compose time

Key Features

• Drag-and-drop ingestion — PDFs, DOCX, Markdown, HTML, GitHub repos
• Pluggable Readers / Chunkers / Embedders / Generators — pick per project
• Multi-LLM — OpenAI, Anthropic, Ollama (local), Cohere, Mistral, OpenRouter
• Citations — answers cite which chunks they came from, with previews
• Conversational memory — multi-turn chat with context
• Hybrid search — vector + keyword via Weaviate's BM25
• Self-hosted Weaviate — embedded mode for zero infra setup
• Export / inspect — see chunks, embeddings, and retrieval results in the UI

Comparison with Similar Tools

FAQ

Q: Verba vs Anything LLM? A: Both are RAG chat apps. Verba is opinionated around Weaviate; AnythingLLM has more vector DB choices and richer multi-workspace features. Verba is closer to a reference implementation; AnythingLLM is more product-like.

Q: Do I need a hosted Weaviate? A: No — Verba embeds Weaviate by default (single-process). For multi-user/production, point at a self-hosted or Weaviate Cloud cluster.

Q: Can it use only local models? A: Yes. Combine Ollama (for generation) + Sentence Transformers (for embedding) and you have a fully local RAG stack with no API keys.

Q: How does it handle large document corpora? A: Bottleneck is Weaviate's capacity. Embedded mode handles tens of thousands of chunks comfortably; for millions, run a real Weaviate cluster.

Sources

• GitHub: https://github.com/weaviate/Verba
• Company: Weaviate
• License: BSD-3-Clause