RAG: Documents & Search.

Drag → Drop → Production RAG

Drag-and-drop RAG builder with automatic evals that find the best configuration for extraction, embeddings, chunking, and indexing.

macOS · Windows · Linux

RAG in Kiln

Build a RAG in minutes

Just drop in your documents, and Kiln takes care of the rest. We extract, chunk, embed, index, eval and optimize.

Find the optimal RAG config

Different data has different RAG needs. Kiln can explore every variable to find what works best for your agent.

Automatic RAG evals

Kiln builds RAG evals automatically, from your docs. It generates synthetic Q&A pairs, and tests your RAG to ensure it can retrieve all relevant data.

Suggested configs get you searching fast

Pick from four ready-made search configurations — Best Quality, Cost Optimized, Vector Only, or OpenAI Based — and start retrieving immediately. Each pre-selects extraction, chunking, embedding, and indexing so you skip setup entirely.

Everything you need for production RAG

Multimodal documents

PDFs, images, video, and audio—not just text files.

Hybrid search

Combine semantic meaning with exact keyword matching via LanceDB.

Custom extraction prompts

Filter irrelevant content from documents for cleaner retrieval.

Configurable chunking

Control chunk size, overlap, and top-K for your retrieval pattern.

Document tagging

Organize docs into sections and target specific sets per search tool.

RAG accuracy evals

Auto-generated Q&A pairs to systematically test retrieval quality.

Extraction reuse

Shared extraction configs skip reprocessing across search tools.

Open source & local-first

Your documents stay on your machine. MIT-licensed Python library, source-available app.

RAG before and after Kiln

Without Kiln
  • Glue together extraction, chunking, embedding, and retrieval across five different libraries and configs.
  • Guess whether a chunking or embedding change actually improved results—or just felt like it did.
  • Re-run the entire pipeline from scratch every time you tweak a setting.
With Kiln
  • Drag in documents, pick a suggested config, and search in under 5 minutes.
  • Measure retrieval quality with auto-generated Q&A evals and systematic comparison.
  • Kiln reuses extraction and embedding work—swap a search strategy without reprocessing.

Frequently asked

What file types does Kiln support for RAG?

Documents (PDF, TXT, MD, HTML), images (JPG, JPEG, PNG), videos (MP4, MOV), and audio (MP3, WAV, OGG). Gemini provides the widest coverage; OpenAI-based extraction does not support audio or video.

Do my documents leave my machine?

Documents stay on your filesystem. Content is sent to AI provider APIs only during extraction and embedding — the same providers you already use.

How does hybrid search work?

It combines vector search (semantic meaning) with full-text search (BM25 keyword matching) via LanceDB. Hybrid is recommended for most use cases; vector-only and full-text-only are also available.

How do I measure retrieval quality?

Kiln generates Q&A pairs from your documents and uses them as an eval dataset. A judge compares responses to known-correct references, avoiding the circular problem of an LLM judging knowledge it lacks.

Ship RAG you can actually measure.

Drag in your documents, optimize with real evals, and deploy to production.