Demonstrates the retrieval-augmented generation pattern using Zerfoo. A hardcoded 5-document corpus is embedded, a query is matched against the corpus via cosine similarity, and the top-3 most relevant documents are passed as context to the model for generation.

Retrieval-Augmented Generation grounds a language model's responses in specific documents rather than relying solely on its training data. The pattern has three steps:

1. Embed — Convert documents and the query into vector embeddings using the model.
2. Retrieve — Find the most relevant documents by comparing embedding similarity.
3. Generate — Pass the retrieved documents as context in the prompt, so the model answers based on the provided facts.

This approach reduces hallucination and lets you inject domain-specific knowledge without fine-tuning.

• Go 1.25+
• A GGUF model file (e.g., Gemma 3 1B or Llama 3.2 1B)

pip install huggingface-hub

huggingface-cli download google/gemma-3-1b-it-qat-q4_0-gguf \
  --local-dir ./models

go build -o rag ./examples/rag/

./rag --model ./models/gemma-3-1b-it-qat-q4_0.gguf

With a custom query:

./rag --model ./models/gemma-3-1b-it-qat-q4_0.gguf \
  --query "What algorithm does Go's garbage collector use?"

Top-3 documents (by similarity):
  1. [0.9234] Go's garbage collector uses a concurrent tri-color mark-and-sweep algorithm.
  2. [0.7812] Go uses goroutines for concurrency, which are lightweight threads managed by the Go runtime.
  3. [0.7456] Go was created at Google in 2009 by Robert Griesemer, Rob Pike, and Ken Thompson.

Go's garbage collector uses a concurrent tri-color mark-and-sweep algorithm...

1. The model is loaded via zerfoo.Load(), which accepts a local GGUF path or a HuggingFace model ID.
2. model.Embed() computes vector embeddings for each corpus document and the query.
3. embedding.CosineSimilarity() ranks corpus documents by relevance to the query.
4. The top-3 documents are injected into a prompt as context, and model.Chat() generates an answer grounded in those facts.

In a production system, the corpus would be stored in a vector database and the embedding step would happen at index time rather than query time.