A minimal example that shows how to build a simple RAG-style workflow in Go using:

• Firebase Genkit for Go (flows, prompts, retrievers)
• Milvus as the vector database
• OpenAI for embeddings and chat completion

The program:

1. Starts a Milvus client and ensures a collection named "products" exists (with a float vector field and text field).
2. Indexes four sample product descriptions using OpenAI embeddings (text-embedding-3-small).
3. Retrieves the top 2 most relevant items for a user query from Milvus.
4. Feeds the retrieved context into a prompt and asks GPT-4o to answer like a phone store salesperson.

• Go 1.25 or newer
• Docker and Docker Compose
• An OpenAI API key (environment variable OPENAI_API_KEY)

Note: This example uses a local replacement for Genkit in go.mod:

replace github.com/firebase/genkit/go v1.0.2 => /Users/gvo/work/genkit-fork/go

If you are not on the same machine or do not have that path, update/remove this replace directive accordingly.

1. Start Milvus locally via Docker Compose (from the project root):

docker compose up -d

This brings up etcd, MinIO, and a Milvus standalone instance. Milvus will be available at 127.0.0.1:19530 when healthy.

2. Set your OpenAI API key:

macOS/Linux:

export OPENAI_API_KEY=your_key_here

Windows (PowerShell):

$Env:OPENAI_API_KEY = "your_key_here"

3. Run the example:

go run .

On the first run, the app will create and load the products collection (if it does not exist), index four tiny sample documents, run a retrieval for the query "I want to buy an iphone", and then call GPT-4o with the retrieved context. You should see a log line like:

result: ...

4. Stop the services when done:

docker compose down

• Milvus connection: main.go connects to Milvus at 127.0.0.1:19530 (see milvus.NewEngine(ctx, milvus.WithAddress("127.0.0.1:19530"))).
• Collection schema (collection name: products):
  • Primary key: id (int64)
  • Vector: vector (FloatVector, dim 1536)
  • Text: text (VarChar)
  • An index is created using cosine similarity.
• Embeddings: OpenAI model text-embedding-3-small is used to embed documents before indexing.
• Retrieval: Uses Genkit's Milvus retriever, requesting top 2 results.
• Prompt: A simple template instructs the assistant to act as a phone store salesperson and use the retrieved context.

Adjust these in main.go if needed:

• Milvus address: 127.0.0.1:19530
• Collection name: products
• Keys and dimensions: id, vector (dim 1536), text
• Embedding model: text-embedding-3-small
• Chat model: openai/gpt-4o
• Retrieval limit: 2

Also ensure the go.mod replace directive for Genkit points to a valid path for your environment, or remove it to use the published module.

• Milvus not ready: If the app fails on first connection, wait until containers are healthy (docker compose ps) and try again.
• Missing OPENAI_API_KEY: Set the environment variable before running.
• Network or model errors: Ensure your network allows calls to OpenAI and that your account has access to the specified models.
• Schema/compatibility: If you change vector dimensions or model, update both the Milvus schema and the embedder dimension accordingly.