README
¶
Azure OpenAI — RAG (Retrieval-Augmented Generation)
End-to-end RAG on Azure OpenAI: embed documents with an Azure embeddings deployment, retrieve the most relevant ones for a question, and feed them as grounding context into an Azure chat completion.
What it demonstrates
model.NewAzureOpenAIEmbeddingsWithConfigas amodel.EmbeddingsProviderknowledge.NewVectorKnowledge—AddDocuments+Loadto index,Searchto retrieve top-k- A self-contained in-memory
storage.VectorStore(MemoryVectorStore) using cosine similarity. Chronos has no built-in in-memory vector store, so this file doubles as a reference implementation of thestorage.VectorStoreinterface (CreateCollection,Upsert,Search,Delete,Close) - Grounded answering: retrieved documents are injected as context with an instruction to answer only from that context
Run
This example needs two deployments — one chat and one embeddings:
export AZURE_OPENAI_API_KEY=<your-azure-api-key>
export AZURE_OPENAI_ENDPOINT=https://<your-resource>.openai.azure.com
export AZURE_OPENAI_DEPLOYMENT=<your-chat-deployment> # e.g. gpt-4o-mini
export AZURE_OPENAI_EMBED_DEPLOYMENT=<your-embeddings-deployment> # e.g. text-embedding-3-small
export AZURE_OPENAI_API_VERSION=2024-10-21
go run ./examples/azure_rag/main.go
embedDimension in main.go is set to 1536 (text-embedding-3-small); change
it to match your embeddings deployment (e.g. 3072 for
text-embedding-3-large).
Without AZURE_OPENAI_API_KEY the example prints the required variables and
exits 0 — no network call.
Tests
main_test.go is fully offline: it tests MemoryVectorStore (upsert/replace,
delete, top-k) and cosineSimilarity ranking with tiny deterministic vectors.
It never contacts Azure.
go test ./examples/azure_rag/...
See also: examples/azure (chat/streaming) and
examples/azure_tools (tool calling).
Documentation
¶
Overview ¶
Azure OpenAI RAG (retrieval-augmented generation) example for Chronos.
This example shows an end-to-end RAG flow on Azure OpenAI:
- Build a small self-contained in-memory VectorStore (cosine similarity). Chronos ships production vector adapters (e.g. Qdrant), but there is no built-in in-memory store — so this file doubles as a reference for how to implement the storage.VectorStore interface.
- Wrap it in a knowledge.VectorKnowledge that embeds documents with an Azure OpenAI embeddings deployment (model.NewAzureOpenAIEmbeddings...).
- Load a few sample documents.
- Answer a question by retrieving the most relevant documents and feeding them as grounding context into an Azure chat completion.
Prerequisites:
- An Azure OpenAI resource with BOTH a chat deployment (e.g. gpt-4o-mini) and an embeddings deployment (e.g. text-embedding-3-small)
- Go 1.24+
Set the following environment variables before running:
export AZURE_OPENAI_API_KEY=<your-azure-api-key> export AZURE_OPENAI_ENDPOINT=https://<your-resource>.openai.azure.com export AZURE_OPENAI_DEPLOYMENT=<your-chat-deployment> # chat export AZURE_OPENAI_EMBED_DEPLOYMENT=<your-embeddings-deployment> # embeddings export AZURE_OPENAI_API_VERSION=2024-10-21
Run:
go run ./examples/azure_rag/main.go
If AZURE_OPENAI_API_KEY is not set the example prints the required variables and exits cleanly (0) without making any network call. The in-memory VectorStore itself is exercised offline by main_test.go.
Source Files