Skip to main content

Chat with your knowledge base

A language model knows what was in its training data and nothing about your policies, contracts, or product manuals. This page explains why retrieval-augmented generation (RAG) is the standard answer, what DIAL provides out of the box, and how to build a grounded assistant over your own documents. It is for evaluators judging whether DIAL fits a knowledge-base scenario, and for builders ready to start.

Why this matters​

Teams reach for "chat with our documents" earlier than almost any other AI scenario, because the value is obvious: support agents answering from the current handbook, engineers querying internal runbooks, analysts asking questions of a contract set. Without grounding, the model invents plausible answers, which is worse than no answer when the subject is a refund policy or a compliance rule.

RAG closes the gap by retrieving relevant passages from a trusted corpus and supplying them as context, so answers cite real, current, organization-specific material. In an enterprise, retrieval is also a governance boundary: what a user can retrieve must respect the same access control as everything else, and the indexed content must live in storage you control.

What you can do with DIAL​

DIAL ships a reference RAG implementation, DIAL RAG, as a first-class application. You attach documents to a chat request and get back a grounded answer. Because it runs as a DIAL application, it reaches models through DIAL Core and inherits the platform's access control, file storage, and observability rather than calling providers directly.

Its retrieval is hybrid and multimodal: it combines semantic, keyword, and image-based retrievers, so it finds answers in prose, exact identifiers, and content that lives in diagrams or scanned pages. Indexing happens once and is cached and reused, so repeat questions against the same corpus skip the expensive parsing step. You can adopt the whole pipeline, or use retrieval-only mode and run your own generation on top.

How you'd build it​

  • Understand the approach first. DIAL RAG explains the document formats, the retrievers, and index reuse.
  • Choose the document lifecycle. Compare DIAL RAG and Generic RAG, then use the Generic RAG guide for a collection uploaded and indexed in advance.
  • Build the app. Tutorial: build a RAG app is the hands-on guided build with a sample repo and data.
  • Handle documents. File management covers uploading and retrieving the files that become your corpus.
  • Support rich documents. Multimodality explains how DIAL handles images, scans, and mixed content end to end.
  • Measure quality. Evaluation in DIAL Admin scores retrieval and answers on your own test cases, so you can tune rather than guess.

Access control is the part teams underestimate. Because DIAL RAG runs behind Core, the same access rules that govern the rest of the platform govern what each user's queries can retrieve, so a grounded assistant does not become a way around document permissions.

Why DIAL over the alternatives​

Frameworks such as LangChain and LlamaIndex give you primitives to assemble a pipeline from scratch, with full control over chunking, retrieval, and reranking. DIAL RAG gives you an assembled, governed pipeline with multimodal retrieval, index caching, and platform integration already done. The trade-off is control versus integration: choose DIAL RAG when you want grounded answers with platform governance and little infrastructure to run, and build your own when your pipeline needs a retrieval strategy DIAL RAG does not express. See DIAL RAG for the retrievers and index reuse.

Further reading​