rag/

Retrieval-augmented generation (RAG)

These examples follow the five .NET AgentWithRAG samples at revision 0c9944c, preserving their source documents, questions, retrieval settings, and context/history behavior.

Configuration

Set the Foundry project endpoint and sign in using a credential supported by DefaultAzureCredential, such as az login:

$env:FOUNDRY_PROJECT_ENDPOINT = "https://<resource>.services.ai.azure.com/api/projects/<project>"
$env:FOUNDRY_MODEL = "gpt-5.4-mini"
$env:FOUNDRY_EMBEDDING_MODEL = "text-embedding-3-large"

The examples use the shared demo.FoundryModel, which defaults to gpt-4o-mini when FOUNDRY_MODEL is unset or blank. Set it as above to match the upstream sample's gpt-5.4-mini choice. Steps 01 and 02 also use FOUNDRY_EMBEDDING_MODEL, which defaults to text-embedding-3-large for their 3072-dimensional vector schemas. Deployment names must match those in your resource.

For production, choose a specific credential appropriate to the deployment, rather than the development-oriented DefaultAzureCredential chain.

Run an example from the repository root:

go run .\examples\02-agents\rag\step01_basic_text_rag
go run .\examples\02-agents\rag\step02_custom_vector_store_rag
go run .\examples\02-agents\rag\step03_custom_rag_data_source
go run .\examples\02-agents\rag\step04_foundry_service_rag
go run .\examples\02-agents\rag\step05_neo4j_graph_rag

These commands invoke live services and may incur charges.

Qdrant

Start a disposable local Qdrant instance with its gRPC port exposed, for example:

docker run --rm -p 6333:6333 -p 6334:6334 qdrant/qdrant

Step 02 deletes and recreates the documentation collection on every run, matching the upstream sample. Do not point it at a collection containing data you want to retain. It downloads the same two Agent Framework documentation pages, splits them into 2000 UTF-16 code units with 200 units of overlap, and stores the text, source metadata, and an unnamed cosine vector in Qdrant.

Foundry service RAG

Step 04 embeds the upstream knowledge-base file, so it works independently of the current working directory.

It creates a file and a vector store, waits up to five minutes for the file to finish indexing, and then creates a version of the agent named AskContoso. Failed or cancelled ingestion stops provisioning and triggers cleanup. It invokes that exact version through the project Responses API using an agent_reference, without overriding server-owned instructions or tools.

The example deletes AskContoso during cleanup, including its versions. Run it only in a project where this name is disposable. Cleanup also removes the uploaded file and vector store. Go defers additionally attempt cleanup after partial failures and cancellation; cleanup failures remain visible.

Neo4j

$env:NEO4J_URI = "neo4j://localhost:7687"
$env:NEO4J_USERNAME = "neo4j"
$env:NEO4J_PASSWORD = "<password>"
$env:NEO4J_FULLTEXT_INDEX_NAME = "search_chunks"

The username and index name default to neo4j and search_chunks. The graph must already contain the same document/company/product/risk schema used by the upstream sample, with a full-text index over the chunk text. The example does not create the graph or seed data. See the upstream Neo4j integration for graph setup guidance.

The provider uses the ten most recent nonblank user/assistant messages, the upstream stop-word list, read-routed Cypher, limits before and after graph enrichment, and the same score/metadata formatting. It does not generate embeddings for this full-text sample. Retrieved graph context is excluded from retained history so later searches use conversation turns rather than prior search results. The driver is closed before reporting operation or cleanup errors.

Go implementation notes

• Each example contains its own instructions, questions, and conversation loop; there is no shared RAG support package. Step 01 keeps its document schema and seed data alongside its entrypoint.
• Examples use the shared demo.NewLogger middleware and demo.Response output. demo.FoundryEmbeddingModel holds the embedding deployment resolved at startup.
• Model-backed agents use foundryprovider.NewAgent. Local conversation history is enabled explicitly with DisableStoreOutput; Steps 01, 02, and 05 exclude retrieved context from retained history.
• Steps 01 and 02 use foundryprovider.NewEmbeddingGenerator for embeddings. It derives the resource OpenAI endpoint from the project URL, configures Foundry authentication, and delegates to the OpenAI embedding adapter.
• The existing project client does not expose file/vector-store operations or agent-version administration. Step 04 uses foundryprovider.NewOpenAIClient for those operations, retaining Azure credential refresh and origin protections.
• Step 04 uses the native SDK's generic HTTP methods for agent administration and an explicit version reference, rather than attaching to a potentially different version selected by an agent endpoint.
• Step 01 uses the SDK's in-memory connector for storage, batch embedding generation, and cosine top-1 search. Its document schema declares a 3072-dimensional vector generated from the text. embedding.Bind connects the Foundry generator to the collection options. UUID keys are generated for each upload; source IDs remain metadata, so repeated uploads create separate records.
• Qdrant and Neo4j use native Go connectors, not mock substitutes.
• Neo4j retrieval helpers follow Neo4j.AgentFramework.GraphRAG 0.1.0-preview.2; its MIT notice is included in LICENSE.neo4j.

Local validation

go build .\examples\02-agents\rag\...
go vet .\examples\02-agents\rag\...

These checks do not require credentials or running Qdrant, Neo4j, or Foundry services.