Every team building with LLMs hits the same wall: the model is smart but knows nothing about your documents. The fix has a name — retrieval-augmented generation — and a reputation for turning a one-week feature into a one-quarter infrastructure project. It doesn't have to.
What RAG actually is
RAG means the model looks things up before it answers. Your documents get split into chunks, each chunk gets embedded into a vector, and at question time the system retrieves the most relevant chunks and hands them to the model as context. The model answers from what it was shown — with citations — instead of guessing from training data. That's the whole idea; the full walkthrough is in our guide to building a RAG pipeline.
Why self-building it drags on
None of the individual pieces is exotic. The drag comes from how many there are and how they interact:
- Ingestion — PDFs, Word files, scans, and web pages all need different parsing before you have clean text. This stage alone sinks most timelines.
- Chunking — split wrong and retrieval returns half-tables and orphaned sentences; every corpus needs tuning.
- Embeddings — pick a model, run it as a service, and commit: change it later and you re-embed everything.
- Retrieval — vector similarity alone misses exact terms, so production systems add keyword search and rank fusion. Two indexes to keep in sync.
- Reranking — the step that makes the top result the right result, and one more model to host.
Each piece is a service to deploy, monitor, and upgrade. The parts are cheap; the plumbing is expensive — and it's all undifferentiated work. Nobody chooses your product because you wrote your own chunker.
The alternative: retrieval behind one API
Kit for AI runs that whole pipeline as a service. You send documents — files, URLs, even YouTube transcripts — into a knowledge base; conversion to clean Markdown, structure-aware chunking, embedding, and indexing happen on ingest. Queries run hybrid semantic-plus-full-text search with reranking, and return cited passages. The retrieval stack is built in-house, not stitched from third-party wrappers, which means one place to send data and one place to ask questions.
Custom doesn't mean self-hosted
The knowledge base is still entirely yours: your documents, your corrections, your project boundaries — spaces keep separate projects' knowledge isolated from each other. "Custom RAG" was never really about owning the vector database; it's about the answers being grounded in your content. That part you keep. The pager duty for six services, you skip.
Two ways to wire it in
Over the REST API, ingest and search are plain endpoints — build them into any app or pipeline. Over MCP, retrieval becomes a tool the agent calls by itself, mid-conversation: it can search your knowledge base, cite what it found, and even add new sources without leaving the chat. The API reference and setup guides are in the docs. Start free with a real document and a real question — you'll know within minutes whether the retrieval holds up.