Impact.Bedrock (Impact v0.0.1)

Copy Markdown

Span helpers for AWS Bedrock Converse-style API calls.

Use these when you need to manually instrument Bedrock calls that don't go through Req (and therefore can't be auto-instrumented by Impact.Instrumentation.Bedrock). Typical case: apps using the :aws_elixir (AWS.BedrockRuntime) or :ex_aws_bedrock clients.

High-level usage — with_call/4

Wraps the entire LLM call in a bedrock_call span. Sets request-side attributes upfront, response-side attributes after the closure returns, and latency on every code path:

Impact.Bedrock.with_call(model, body, [conversation_id: agent_id], fn ->
  case AWS.BedrockRuntime.converse(client, model.arn, body, opts) do
    {:ok, response, _metadata} -> {:ok, response}
    {:error, reason} -> {:error, reason}
  end
end)

Returns whatever the closure returned, unchanged.

Result classification

The closure can return any of these shapes; with_call/4 interprets them as follows:

Return valueAttribute setLatency stamped
{:ok, response}attrs_for_response/2yes
{:ok, response, _meta} (3-tuple aws-elixir form)attrs_for_response/2yes
{:error, reason}attrs_for_error/2yes
anything else (e.g. {:snooze, n})aws.bedrock.latency_ms onlyyes

Manual control — attrs_for_request/3 / attrs_for_response/2 / attrs_for_error/2

Build attribute maps directly when you need to drive the trace block yourself (multi-attempt retries, snooze handling with per-attempt timing, etc):

request_attrs = Impact.Bedrock.attrs_for_request(model, body, conversation_id: agent_id)

Impact.trace [type: :llm, name: "bedrock_call", attributes: request_attrs] do
  start_ms = System.monotonic_time(:millisecond)
  # ... your call ...
  Impact.set_attributes(Impact.Bedrock.attrs_for_response(response, latency))
end

Attribute keys

Request-side (attrs_for_request/3):

  • gen_ai.system = "aws.bedrock"
  • gen_ai.provider.name = "aws"
  • gen_ai.operation.name = "chat"
  • gen_ai.request.model — model name (from model.name or ARN)
  • gen_ai.input.messages — JSON-encoded request messages
  • gen_ai.system_instructions — extracted system text
  • gen_ai.conversation.id — opt :conversation_id
  • gen_ai.model.id — opt :model_id
  • aws.bedrock.model.arn — from model.arn
  • aws.bedrock.inference_config — JSON-encoded inferenceConfig
  • aws.bedrock.tool_config — JSON-encoded toolConfig

Response-side (attrs_for_response/2):

  • gen_ai.response.id
  • gen_ai.response.finish_reasons — list with stopReason
  • gen_ai.usage.input_tokens, gen_ai.usage.output_tokens
  • gen_ai.output.messages — JSON-encoded assistant message
  • gen_ai.output.text — extracted plain text
  • gen_ai.response.tool_calls.count — number of toolUse blocks
  • aws.bedrock.latency_ms

Error-side (attrs_for_error/2):

  • error.message — inspect(reason)
  • aws.bedrock.latency_ms

For tool / function call spans, see Impact.Tool.

Summary

Functions

attrs_for_error(reason, latency_ms)

Build error-side attributes for a failed bedrock_call.

attrs_for_request(model, body, opts \\ [])

Build request-side attributes for a bedrock_call span.

attrs_for_response(response, latency_ms)

Build response-side attributes for a successful bedrock_call.

with_call(model, body, opts \\ [], fun)

Wrap a Bedrock-style API call in a bedrock_call LLM span. Handles request attrs, response/error attrs, and latency on every code path.

Functions

attrs_for_error(reason, latency_ms)

@spec attrs_for_error(any(), integer()) :: map()

Build error-side attributes for a failed bedrock_call.

attrs_for_request(model, body, opts \\ [])

@spec attrs_for_request(map(), map(), keyword()) :: map()

Build request-side attributes for a bedrock_call span.

Options

  • :conversation_id — value for gen_ai.conversation.id
  • :model_id — value for gen_ai.model.id (internal model identifier distinct from the human-readable model name)

attrs_for_response(response, latency_ms)

@spec attrs_for_response(map(), integer()) :: map()

Build response-side attributes for a successful bedrock_call.

with_call(model, body, opts \\ [], fun)

@spec with_call(map(), map(), keyword(), (-> any())) :: any()

Wrap a Bedrock-style API call in a bedrock_call LLM span. Handles request attrs, response/error attrs, and latency on every code path.

Options (forwarded to attrs_for_request/3)

  • :conversation_id — value for gen_ai.conversation.id
  • :model_id — value for gen_ai.model.id