ChromaDB.SemanticKernel.Sample

Semantic Kernel samples that use Chroma as the vector store through ChromaDotNet.VectorData.

They follow the structure of the Semantic Kernel .NET samples: the concept samples and the text search steps. Each sample is an xUnit test that runs against real models when it needs one. Chroma replaces the vector store that the original sample uses. The samples that use a chat model also come with models that run locally in Ollama.

This is a community project. It is not affiliated with or endorsed by Chroma.

Website: chromadotnet.org

Samples

Sample	Description
VectorStore_VectorSearch_MultiStore_Chroma	Ingests a glossary into Chroma and searches it, with and without a filter. It goes through the common vector store sample code, with and without dependency injection. Adapted from VectorStore_VectorSearch_MultiStore_Qdrant.
VectorStore_DynamicDataModel_Interop_Chroma	Writes records as dictionaries described by a record definition, then reads them back as a .NET data model. It also does the reverse. Adapted from VectorStore_DynamicDataModel_Interop.
VectorStore_VectorSearch_Paging_Chroma	Uses Top and Skip to page through vector search results over 1,000 records. Needs no model. Adapted from VectorStore_VectorSearch_Paging, but with the Euclidean distance. Chroma searches with an approximate index, and with the cosine distance that index can miss some of the made-up vectors in the sample.
VectorStore_HybridSearch_Simple_Chroma	Searches a glossary on Chroma Cloud with a vector and keywords, with and without a filter. On Chroma Cloud, the collection gets a BM25 index for the full-text indexed property when it is created. Adapted from VectorStore_HybridSearch_Simple_AzureAISearch.
ChatCompletion_Rag_Chroma	A ChatCompletionAgent that answers through a TextSearchProvider, from documents that a TextSearchStore keeps in Chroma. It also covers citations and a search namespace. Adapted from ChatCompletion_Rag.
ChatCompletion_Rag_Chroma_Ollama	The same sample with models that run locally in Ollama.
Step5_Search_With_Chroma	Searches records that have their own data model in Chroma, through VectorStoreTextSearch. It then either passes the results to the model in a Handlebars prompt, or gives the model the search as a function to call. Adapted from Step4_Search_With_VectorStore.
Step5_Search_With_Chroma_Ollama	The same sample with models that run locally in Ollama.

The samples need the .NET 10 SDK. They also need Docker, except the hybrid search sample, which runs on Chroma Cloud. Their fixtures start Chroma in a container and remove it at the end.

• The fixture for the concept samples starts Chroma on port 8000, as the Semantic Kernel vector store samples do with their databases. Those samples therefore also need port 8000 to be free.
• The fixture for the text search steps starts Chroma on a free port with ChromaDotNet.Testcontainers.

Run the samples with Azure OpenAI

The samples read the same settings as the Semantic Kernel samples. You can set them in user secrets or in environment variables such as AzureOpenAI__Endpoint. The two projects read the same user secrets:

dotnet user-secrets set "AzureOpenAI:Endpoint" "https://your-resource.openai.azure.com/" --project samples/Concepts
dotnet user-secrets set "AzureOpenAI:ChatDeploymentName" "..." --project samples/Concepts

dotnet user-secrets set "AzureOpenAIEmbeddings:Endpoint" "https://your-resource.openai.azure.com/" --project samples/Concepts
dotnet user-secrets set "AzureOpenAIEmbeddings:DeploymentName" "..." --project samples/Concepts

The samples ask the embedding deployment for 1,536 dimensions. The deployment must use a model that can produce them, such as text-embedding-3-small or text-embedding-3-large.

The samples sign in with the Azure CLI (az login). The identity must be able to use the deployments. For example, it can have the Cognitive Services OpenAI User role, or the Foundry User role on a Microsoft Foundry resource.

dotnet test samples/Concepts --filter "FullyQualifiedName~Memory.VectorStore_VectorSearch_MultiStore_Chroma." --logger "console;verbosity=detailed"
dotnet test samples/Concepts --filter "FullyQualifiedName~Memory.VectorStore_DynamicDataModel_Interop_Chroma." --logger "console;verbosity=detailed"
dotnet test samples/Concepts --filter "FullyQualifiedName~Agents.ChatCompletion_Rag_Chroma." --logger "console;verbosity=detailed"
dotnet test samples/GettingStartedWithTextSearch --filter "FullyQualifiedName~Step5_Search_With_Chroma." --logger "console;verbosity=detailed"

The hybrid search sample runs on Chroma Cloud, which supports hybrid search. It uses the tenant and the database from the Chroma Cloud dashboard, and it deletes its collection at the end:

dotnet user-secrets set "Chroma:Endpoint" "https://api.trychroma.com" --project samples/Concepts
dotnet user-secrets set "Chroma:ApiKey" "..." --project samples/Concepts
dotnet user-secrets set "Chroma:Tenant" "..." --project samples/Concepts
dotnet user-secrets set "Chroma:Database" "..." --project samples/Concepts

dotnet test samples/Concepts --filter "FullyQualifiedName~Memory.VectorStore_HybridSearch_Simple_Chroma." --logger "console;verbosity=detailed"

The paging sample needs no model:

dotnet test samples/Concepts --filter "FullyQualifiedName~Memory.VectorStore_VectorSearch_Paging_Chroma." --logger "console;verbosity=detailed"

Run the samples with Ollama

compose.yaml runs Ollama and downloads qwen2.5:3b and nomic-embed-text. The first time, that is about 2 GB. To follow the progress, run docker compose logs -f ollama-models.

docker compose up -d
dotnet test samples/Concepts --filter "FullyQualifiedName~Agents.ChatCompletion_Rag_Chroma_Ollama." --logger "console;verbosity=detailed"
dotnet test samples/GettingStartedWithTextSearch --filter "FullyQualifiedName~Step5_Search_With_Chroma_Ollama." --logger "console;verbosity=detailed"

The defaults match the compose file. To change them, set:

• Ollama:Endpoint
• Ollama:ModelId
• Ollama:EmbeddingModelId
• Ollama:EmbeddingDimensions

The data model of Step5_Search_With_Chroma_Ollama has the 768 dimensions of nomic-embed-text in its attribute. You can find it in ChromaVectorStoreOllamaFixture.cs.

A larger chat model than qwen2.5:3b gives better answers. qwen2.5:3b answers from the documents, but it can add made-up links when they have no source.

Tests

dotnet test test/ChromaDB.SemanticKernel.Sample.Tests

GitHub Actions runs them on every push to main and on every pull request to main.

The tests run the scenario of each sample with the same configuration, except the hybrid search, which needs Chroma Cloud. They run against Chroma in a container that ChromaDotNet.Testcontainers starts on a free port. They need Docker but no model. The embeddings come from word hashes, and the chat model only records what it receives.

• Vector search: each search finds the glossary entry it asks about. The category filter leaves out the entry in the other category, Connectors. Both checks pass with and without dependency injection.
• Data models: a record written as a dictionary reads back as the .NET data model, and the other way round, including the vector.
• Paging: the pages go through each of the 1,000 records once, ordered by distance.
• RAG: the document that answers each question reaches the model. With a search namespace, only the documents in that namespace reach it, along with their source.
• Text search: the search returns the records with their key, text and link. The Handlebars prompt receives the search results. With function calling, the search results reach the model.

The following files come from the Semantic Kernel repository (MIT):

• the files in samples/InternalUtilities
• the common vector store sample code
• the container helper

They are reduced to what these samples use and keep their copyright notice. The samples and their fixtures are adaptations of the Semantic Kernel samples linked in the table above. They are under the same license and carry the same copyright notice.