appsettings.json

{
  "KernelMemory": {
    "Services": {
      "AzureAIContentSafety": {
        // "ApiKey" or "AzureIdentity". For other options see <AzureAIContentSafetyConfig>.
        // AzureIdentity: use automatic Entra (AAD) authentication mechanism.
        //   When the service is on sovereign clouds you can use the AZURE_AUTHORITY_HOST env var to
        //   set the authority host. See https://learn.microsoft.com/dotnet/api/overview/azure/identity-readme
        //   You can test locally using the AZURE_TENANT_ID, AZURE_CLIENT_ID, AZURE_CLIENT_SECRET env vars.
        "Auth": "AzureIdentity",
        "Endpoint": "https://<...>",
        "APIKey": "",
        "GlobalSafetyThreshold": 0.0,
        "IgnoredWords": []
      },
      "AzureAIDocIntel": {
        // "APIKey" or "AzureIdentity".
        // AzureIdentity: use automatic Entra (AAD) authentication mechanism.
        //   When the service is on sovereign clouds you can use the AZURE_AUTHORITY_HOST env var to
        //   set the authority host. See https://learn.microsoft.com/dotnet/api/overview/azure/identity-readme
        //   You can test locally using the AZURE_TENANT_ID, AZURE_CLIENT_ID, AZURE_CLIENT_SECRET env vars.
        "Auth": "AzureIdentity",
        // Optional when Auth == AzureIdentity. Leave it null to use the default.
        // When the service is on sovereign clouds, this setting might be necessary to configure Entra auth tokens.
        // See https://github.com/Azure/azure-sdk-for-net/blob/main/sdk/formrecognizer/Azure.AI.FormRecognizer/src/DocumentAnalysisAudience.cs
        "AzureIdentityAudience": null,
        // Required when Auth == APIKey
        "APIKey": "",
        "Endpoint": ""
      },
      "AzureAISearch": {
        // "ApiKey" or "AzureIdentity". For other options see <AzureAISearchConfig>.
        // AzureIdentity: use automatic Entra (AAD) authentication mechanism.
        //   When the service is on sovereign clouds you can use the AZURE_AUTHORITY_HOST env var to
        //   set the authority host. See https://learn.microsoft.com/dotnet/api/overview/azure/identity-readme
        //   You can test locally using the AZURE_TENANT_ID, AZURE_CLIENT_ID, AZURE_CLIENT_SECRET env vars.
        "Auth": "AzureIdentity",
        // Optional when Auth == AzureIdentity. Leave it null to use the default.
        // When the service is on sovereign clouds, this setting might be necessary to configure Entra auth tokens.
        // See https://github.com/Azure/azure-sdk-for-net/blob/main/sdk/search/Azure.Search.Documents/src/SearchAudience.cs
        "AzureIdentityAudience": null,
        // Required when Auth == APIKey
        "APIKey": "",
        "Endpoint": "https://<...>",
        // Hybrid search is not enabled by default. Note that when using hybrid search
        // relevance scores are different, usually lower, than when using just vector search
        "UseHybridSearch": false,
        // Helps improve relevance score consistency for search services with multiple replicas by
        // attempting to route a given request to the same replica for that session. Use this when
        // favoring consistent scoring over lower latency. Can adversely affect performance.
        //
        // Whether to use sticky sessions, which can help getting more consistent results.
        // When using sticky sessions, a best-effort attempt will be made to target the same replica set.
        // Be wary that reusing the same replica repeatedly can interfere with the load balancing of
        // the requests across replicas and adversely affect the performance of the search service.
        //
        // See https://learn.microsoft.com/rest/api/searchservice/documents/search-post?view=rest-searchservice-2024-07-01&tabs=HTTP#request-body
        "UseStickySessions": false
      },
      "AzureBlobs": {
        // "ConnectionString" or "AzureIdentity". For other options see <AzureBlobConfig>.
        // AzureIdentity: use automatic Entra (AAD) authentication mechanism.
        //   When the service is on sovereign clouds you can use the AZURE_AUTHORITY_HOST env var to
        //   set the authority host. See https://learn.microsoft.com/dotnet/api/overview/azure/identity-readme
        //   You can test locally using the AZURE_TENANT_ID, AZURE_CLIENT_ID, AZURE_CLIENT_SECRET env vars.
        "Auth": "AzureIdentity",
        // Optional when Auth == AzureIdentity. Leave it null to use the default.
        // When the service is on sovereign clouds, this setting might be necessary to configure Entra auth tokens.
        "AzureIdentityAudience": null,
        // Azure Storage account name, required when using AzureIdentity auth
        // Note: you can use an env var 'KernelMemory__Services__AzureBlobs__Account' to set this
        "Account": "",
        // Container where to create directories and upload files
        "Container": "smemory",
        // Required when Auth == ConnectionString
        // Note: you can use an env var 'KernelMemory__Services__AzureBlobs__ConnectionString' to set this
        "ConnectionString": "",
        // Setting used only for country clouds
        "EndpointSuffix": "core.windows.net"
      },
      "AzureOpenAIEmbedding": {
        // "ApiKey" or "AzureIdentity"
        // AzureIdentity: use automatic Entra (AAD) authentication mechanism.
        //   You can test locally using the AZURE_TENANT_ID, AZURE_CLIENT_ID, AZURE_CLIENT_SECRET env vars.
        "Auth": "AzureIdentity",
        // Optional when Auth == AzureIdentity. Leave it null to use the default.
        // in which case use this to change the client audience.
        "AzureIdentityAudience": null,
        "Endpoint": "https://<...>.openai.azure.com/",
        "APIKey": "",
        // Your Azure Deployment name
        "Deployment": "",
        // The max number of tokens supported by model deployed
        // See https://learn.microsoft.com/azure/ai-services/openai/concepts/models
        "MaxTokenTotal": 8191,
        // Which tokenizer to use to correctly measure the size of chunks.
        // Supported values: "p50k", "cl100k", "o200k". Leave it empty if unsure.
        // - Use p50k for the old text-davinci-003 models
        // - Use cl100k for the old gpt-3.4 and gpt-4 family, and for text embedding models
        // - Use o200k for the most recent gpt-4o family
        "Tokenizer": "cl100k",
        // The number of dimensions output embeddings should have.
        // Only supported in "text-embedding-3" and later models developed with
        // MRL, see https://arxiv.org/abs/2205.13147
        "EmbeddingDimensions": null,
        // How many embeddings to calculate in parallel. The max value depends on
        // the model and deployment in use.
        // See https://learn.microsoft.com/azure/ai-services/openai/reference#embeddings
        "MaxEmbeddingBatchSize": 1,
        // How many times to retry in case of throttling.
        "MaxRetries": 10,
        // Thumbprints of certificates that should be trusted for HTTPS requests when SSL policy errors are detected.
        // This should only be used for local development when using a proxy to call the OpenAI endpoints.
        "TrustedCertificateThumbprints": []
      },
      "AzureOpenAIText": {
        // "ApiKey" or "AzureIdentity"
        // AzureIdentity: use automatic Entra (AAD) authentication mechanism.
        //   You can test locally using the AZURE_TENANT_ID, AZURE_CLIENT_ID, AZURE_CLIENT_SECRET env vars.
        "Auth": "AzureIdentity",
        // Optional when Auth == AzureIdentity. Leave it null to use the default.
        // in which case use this to change the client audience.
        "AzureIdentityAudience": null,
        "Endpoint": "https://<...>.openai.azure.com/",
        "APIKey": "",
        "Deployment": "",
        // The max number of tokens supported by model deployed
        // See https://learn.microsoft.com/azure/ai-services/openai/concepts/models
        "MaxTokenTotal": 16384,
        // Which tokenizer to use to correctly measure the size of chunks.
        // Supported values: "p50k", "cl100k", "o200k". Leave it empty if unsure.
        // - Use p50k for the old text-davinci-003 models
        // - Use cl100k for the old gpt-3.4 and gpt-4 family, and for text embedding models
        // - Use o200k for the most recent gpt-4o family
        "Tokenizer": "o200k",
        // "ChatCompletion" or "TextCompletion"
        "APIType": "ChatCompletion",
        // How many times to retry in case of throttling.
        "MaxRetries": 10,
        // Thumbprints of certificates that should be trusted for HTTPS requests when SSL policy errors are detected.
        // This should only be used for local development when using a proxy to call the OpenAI endpoints.
        "TrustedCertificateThumbprints": []
      },
      "OpenAI": {
        // Name of the model used to generate text (text completion or chat completion)
        "TextModel": "gpt-4o-mini",
        // The max number of tokens supported by the text model.
        "TextModelMaxTokenTotal": 16384,
        // Supported values: "p50k", "cl100k", "o200k". Leave it empty for autodetect.
        "TextModelTokenizer": "",
        // What type of text generation, by default autodetect using the model name.
        // Possible values: "Auto", "TextCompletion", "Chat"
        "TextGenerationType": "Auto",
        // Name of the model used to generate text embeddings
        "EmbeddingModel": "text-embedding-ada-002",
        // The max number of tokens supported by the embedding model
        // See https://platform.openai.com/docs/guides/embeddings/what-are-embeddings
        "EmbeddingModelMaxTokenTotal": 8191,
        // Supported values: "p50k", "cl100k", "o200k". Leave it empty for autodetect.
        "EmbeddingModelTokenizer": "",
        // OpenAI API Key
        "APIKey": "",
        // OpenAI Organization ID (usually empty, unless you have multiple accounts on different orgs)
        "OrgId": "",
        // Endpoint to use. By default the system uses 'https://api.openai.com/v1'.
        // Change this to use proxies or services compatible with OpenAI HTTP protocol like LM Studio.
        "Endpoint": "",
        // How many times to retry in case of throttling
        "MaxRetries": 10,
        // The number of dimensions output embeddings should have.
        // Only supported in "text-embedding-3" and later models developed with
        // MRL, see https://arxiv.org/abs/2205.13147
        "EmbeddingDimensions": null,
        // How many embeddings to calculate in parallel.
        // See https://platform.openai.com/docs/api-reference/embeddings/create
        "MaxEmbeddingBatchSize": 100
      },
      "AWSS3": {
        "Auth": "AccessKey",
        // AccessKey ID, required when using AccessKey auth
        // Note: you can use an env var 'KernelMemory__Services__AWSS3__AccessKey' to set this
        "AccessKey": "",
        // SecretAccessKey, required when using AccessKey auth
        // Note: you can use an env var 'KernelMemory__Services__AWSS3__SecretAccessKey' to set this
        "SecretAccessKey": "",
        // Required bucket name where to create directories and upload files.
        // Note: you can use an env var 'KernelMemory__Services__AWSS3__BucketName' to set this
        "BucketName": ""
        // Allows to specify a custom AWS or a compatible endpoint
        // Examples: "https://s3.amazonaws.com", "https://s3.us-west-2.amazonaws.com", "http://127.0.0.1:9444"
        // Note: you can use an env var 'KernelMemory__Services__AWSS3__Endpoint' to set this
        // Note: you can test locally using S3 Ninja https://s3ninja.net
        // "Endpoint": "https://s3.amazonaws.com"
      },
      "LlamaSharp": {
        "TextModel": {
          // path to file, e.g. "llama-2-7b-chat.Q6_K.gguf"
          "ModelPath": "",
          // Max number of tokens supported by the model
          "MaxTokenTotal": 4096
          // Optional parameters
          // "GpuLayerCount": 32,
        },
        "EmbeddingModel": {
          // path to file, e.g. "nomic-embed-text-v1.5.Q8_0.gguf"
          "ModelPath": "",
          // Max number of tokens supported by the model
          "MaxTokenTotal": 4096
          // Optional parameters
          // "GpuLayerCount": 32,
        }
      }
    },
    "DataIngestion": {
      // How many memory DB records to insert at once when extracting memories from
      // uploaded documents (used only if the Memory Db supports batching).
      "MemoryDbUpsertBatchSize": 100
    },
    "Retrieval": {
      "SearchClient": {
        // Maximum number of tokens accepted by the LLM used to generate answers.
        // The number includes the tokens used for the answer, e.g. when using
        // GPT4-32k, set this number to 32768.
        // If the value is not set or less than one, SearchClient will use the
        // max amount of tokens supported by the model in use.
        "MaxAskPromptSize": -1,
        // Maximum number of relevant sources to consider when generating an answer.
        // The value is also used as the max number of results returned by SearchAsync
        // when passing a limit less or equal to zero.
        "MaxMatchesCount": 100,
        // How many tokens to reserve for the answer generated by the LLM.
        // E.g. if the LLM supports max 4000 tokens, and AnswerTokens is 300, then
        // the prompt sent to LLM will contain max 3700 tokens, composed by
        // prompt + question + grounding information retrieved from memory.
        "AnswerTokens": 300,
        // Text to return when the LLM cannot produce an answer.
        "EmptyAnswer": "INFO NOT FOUND"
      }
    }
  }
}