This version is still in development and is not considered stable yet. For the latest snapshot version, please use Spring AI 1.0.0-SNAPSHOT!

Observability

Spring AI builds upon the observability features in the Spring ecosystem to provide insights into AI-related operations. Spring AI provides metrics and tracing capabilities for its core components: ChatClient (including Advisor), ChatModel, EmbeddingModel, ImageModel, and VectorStore.

Low cardinality keys will be added to metrics and traces, while high cardinality keys will only be added to traces.

Chat Client

The spring.ai.chat.client observations are recorded when a ChatClient call() or stream() operations are invoked. They measure the time spent performing the invocation and propagate the related tracing information.

Table 1. Low Cardinality Keys
Name Description

gen_ai.operation.name

Always framework.

gen_ai.system

Always spring_ai.

spring.ai.chat.client.stream

Is the chat model response a stream - true or false

spring.ai.kind

The kind of framework API in Spring AI: chat_client.
Table 2. High Cardinality Keys
Name Description

spring.ai.chat.client.advisor.params

Map of advisor parameters.

spring.ai.chat.client.advisors

List of configured chat client advisors.

spring.ai.chat.client.system.params

Chat client system parameters. Optional.

spring.ai.chat.client.system.text

Chat client system text. Optional.

spring.ai.chat.client.tool.function.names

Enabled tool function names.

spring.ai.chat.client.tool.function.callbacks

List of configured chat client function callbacks.

spring.ai.chat.client.user.params

Chat client user parameters. Optional.

spring.ai.chat.client.user.text

Chat client user text. Optional.

Input Data

The ChatClient input data is typically big and possibly containing sensitive information. For those reasons, it is not exported by default.

Spring AI supports exporting input data as span attributes across all tracing backends.

Property Description Default

spring.ai.chat.client.observations.include-input

Whether to include the input content in the observations.

false
If you enable the inclusion of the input content in the observations, there’s a risk of exposing sensitive or private information. Please, be careful!

Chat Client Advisors

The spring.ai.advisor observations are recorded when a call or stream around advisors is performed. They measure the time spent in the advisor (including the time spend on the inner advisors) and propagate the related tracing information.

Table 3. Low Cardinality Keys
Name Description

gen_ai.operation.name

Always framework.

gen_ai.system

Always spring_ai.

spring.ai.advisor.type

Where the advisor applies it’s logic in the request processing, one of BEFORE, AFTER, or AROUND.

spring.ai.kind

The kind of framework API in Spring AI: advisor.
Table 4. High Cardinality Keys
Name Description

spring.ai.advisor.name

Name of the advisor.

spring.ai.advisor.order

Advisor order in the advisor chain.

Chat Model

Observability features are currently supported only for ChatModel implementations from the following AI model providers: Anthropic, Azure OpenAI, Mistral AI, Ollama, OpenAI, Vertex AI, MiniMax, Moonshot, QianFan, Zhiu AI. Additional AI model providers will be supported in a future release.

The gen_ai.client.operation observations are recorded when calling the ChatModel call or stream methods. They measure the time spent on method completion and propagate the related tracing information.

The gen_ai.client.token.usage metrics measures number of input and output tokens used by a single model call.
Table 5. Low Cardinality Keys
Name Description

gen_ai.operation.name

The name of the operation being performed.

gen_ai.system

The model provider as identified by the client instrumentation.

gen_ai.request.model

The name of the model a request is being made to.

gen_ai.response.model

The name of the model that generated the response.
Table 6. High Cardinality Keys
Name Description

gen_ai.request.frequency_penalty

The frequency penalty setting for the model request.

gen_ai.request.max_tokens

The maximum number of tokens the model generates for a request.

gen_ai.request.presence_penalty

The presence penalty setting for the model request.

gen_ai.request.stop_sequences

List of sequences that the model will use to stop generating further tokens.

gen_ai.request.temperature

The temperature setting for the model request.

gen_ai.request.top_k

The top_k sampling setting for the model request.

gen_ai.request.top_p

The top_p sampling setting for the model request.

gen_ai.response.finish_reasons

Reasons the model stopped generating tokens, corresponding to each generation received.

gen_ai.response.id

The unique identifier for the AI response.

gen_ai.usage.input_tokens

The number of tokens used in the model input (prompt).

gen_ai.usage.output_tokens

The number of tokens used in the model output (completion).

gen_ai.usage.total_tokens

The total number of tokens used in the model exchange.

gen_ai.prompt

The full prompt sent to the model. Optional.

gen_ai.completion

The full response received from the model. Optional.
For measuring user tokens, the previous table lists the values present in an observation trace. Use the metric name gen_ai.client.token.usage that is provided by the ChatModel.
Table 7. Events
Name Description

gen_ai.content.prompt

Event including the content of the chat prompt. Optional.

gen_ai.content.completion

Event including the content of the chat completion. Optional.

Chat Prompt and Completion Data

The chat prompt and completion data is typically big and possibly containing sensitive information. For those reasons, it is not exported by default.

Spring AI supports exporting chat prompt and completion data as span events if you use an OpenTelemetry tracing backend, whereas data is exported as span attributes if you use an OpenZipkin tracing backend.

Furthermore, Spring AI supports logging chat prompt and completion data, useful for troubleshooting scenarios.

Property Description Default

spring.ai.chat.observations.include-prompt

Include the prompt content in observations. true or false

false

spring.ai.chat.observations.include-completion

Include the completion content in observations. true or false

false

spring.ai.chat.observations.include-error-logging

Include error logging in observations. true or false

false
If you enable the inclusion of the chat prompt and completion data in the observations, there’s a risk of exposing sensitive or private information. Please, be careful!

EmbeddingModel

Observability features are currently supported only for EmbeddingModel implementations from the following AI model providers: Azure OpenAI, Mistral AI, Ollama, and OpenAI. Additional AI model providers will be supported in a future release.

The gen_ai.client.operation observations are recorded on embedding model method calls. They measure the time spent on method completion and propagate the related tracing information.

The gen_ai.client.token.usage metrics measures number of input and output tokens used by a single model call.
Table 8. Low Cardinality Keys
Name Description

gen_ai.operation.name

The name of the operation being performed.

gen_ai.system

The model provider as identified by the client instrumentation.

gen_ai.request.model

The name of the model a request is being made to.

gen_ai.response.model

The name of the model that generated the response.
Table 9. High Cardinality Keys
Name Description

gen_ai.request.embedding.dimensions

The number of dimensions the resulting output embeddings have.

gen_ai.usage.input_tokens

The number of tokens used in the model input.

gen_ai.usage.total_tokens

The total number of tokens used in the model exchange.
For measuring user tokens, the previous table lists the values present in an observation trace. Use the metric name gen_ai.client.token.usage that is provided by the EmbeddingModel.

Image Model

Observability features are currently supported only for ImageModel implementations from the following AI model providers: OpenAI. Additional AI model providers will be supported in a future release.

The gen_ai.client.operation observations are recorded on image model method calls. They measure the time spent on method completion and propagate the related tracing information.

The gen_ai.client.token.usage metrics measures number of input and output tokens used by a single model call.
Table 10. Low Cardinality Keys
Name Description

gen_ai.operation.name

The name of the operation being performed.

gen_ai.system

The model provider as identified by the client instrumentation.

gen_ai.request.model

The name of the model a request is being made to.
Table 11. High Cardinality Keys
Name Description

gen_ai.request.image.response_format

The format in which the generated image is returned.

gen_ai.request.image.size

The size of the image to generate.

gen_ai.request.image.style

The style of the image to generate.

gen_ai.response.id

The unique identifier for the AI response.

gen_ai.response.model

The name of the model that generated the response.

gen_ai.usage.input_tokens

The number of tokens used in the model input (prompt).

gen_ai.usage.output_tokens

The number of tokens used in the model output (generation).

gen_ai.usage.total_tokens

The total number of tokens used in the model exchange.

gen_ai.prompt

The full prompt sent to the model. Optional.
For measuring user tokens, the previous table lists the values present in an observation trace. Use the metric name gen_ai.client.token.usage that is provided by the ImageModel.
Table 12. Events
Name Description

gen_ai.content.prompt

Event including the content of the image prompt. Optional.

Image Prompt Data

The image prompt data is typically big and possibly containing sensitive information. For those reasons, it is not exported by default.

Spring AI supports exporting image prompt data as span events if you use an OpenTelemetry tracing backend, whereas data is exported as span attributes if you use an OpenZipkin tracing backend.

Property Description Default

spring.ai.image.observations.include-prompt

true or false

false
If you enable the inclusion of the image prompt data in the observations, there’s a risk of exposing sensitive or private information. Please, be careful!

Vector Stores

All vector store implementations in Spring AI are instrumented to provide metrics and distributed tracing data through Micrometer.

The db.vector.client.operation observations are recorded when interacting with the Vector Store. They measure the time spent on the query, add and remove operations and propagate the related tracing information.

Table 13. Low Cardinality Keys
Name Description

db.operation.name

The name of the operation or command being executed. One of add, delete, or query.

db.system

The database management system (DBMS) product as identified by the client instrumentation. One of pg_vector, azure, cassandra, chroma, elasticsearch, milvus, neo4j, opensearch, qdrant, redis, typesense, weaviate, pinecone, oracle, mongodb, gemfire, hana, simple.

spring.ai.kind

The kind of framework API in Spring AI: vector_store.
Table 14. High Cardinality Keys
Name Description

db.collection.name

The name of a collection (table, container) within the database.

db.namespace

The name of the database, fully qualified within the server address and port.

db.record.id

The record identifier if present.

db.search.similarity_metric

The metric used in similarity search.

db.vector.dimension_count

The dimension of the vector.

db.vector.field_name

The name field as of the vector (e.g. a field name).

db.vector.query.content

The content of the search query being executed.

db.vector.query.filter

The metadata filters used in the search query.

db.vector.query.response.documents

Returned documents from a similarity search query. Optional.

db.vector.query.similarity_threshold

Similarity threshold that accepts all search scores. A threshold value of 0.0 means any similarity is accepted or disable the similarity threshold filtering. A threshold value of 1.0 means an exact match is required.

db.vector.query.top_k

The top-k most similar vectors returned by a query.
Table 15. Events
Name Description

db.vector.content.query.response

Event including the vector search response data. Optional.

Response Data

The vector search response data is typically big and possibly containing sensitive information. For those reasons, it is not exported by default.

Spring AI supports exporting vector search response data as span events if you use an OpenTelemetry tracing backend, whereas data is exported as span attributes if you use an OpenZipkin tracing backend.

Property Description Default

spring.ai.vectorstore.observations.include-query-response

true or false

false
If you enable the inclusion of the vector search response data in the observations, there’s a risk of exposing sensitive or private information. Please, be careful!