Skip to main content
Choose what Impact sends from the application. The SDK observes semantic provider data as the native client exposes it and applies its policy to the Impact telemetry copy.

Choose what Impact captures

Content capture is on by default. Use the initialization policy to omit selected categories from Impact telemetry.
Every omitted setting defaults to true. input includes instructions and requested schemas. tools covers arguments and results on tool carriers, retrieval includes memory content, and media controls observed bytes. The SDK validates and snapshots configuration before activation. Mixed native blocks retain permitted content. If the SDK can’t redact the affected Impact record safely, it drops that record and reports the reason. The root span or log records the effective policy and omission markers. If the app returns a tool result as its own request output, that copy follows the output policy. Restrict output too if it can contain data you want omitted. Vercel AI SDK provider metadata can mix several categories in one opaque value. Impact captures that value only when all five categories permit capture and the native telemetry settings permit outputs. Native per-call recordOutputs: false or disabled telemetry takes precedence. Only the Impact copy is projected. Existing customer records, processors and destinations remain untouched. Under a restriction, unclassified correlated log bodies, error text and free-form feedback are omitted because they can quote any content category. Numeric feedback and operation identity remain available. Simulation invocation and Protect inference are separately authorized application work. Their telemetry copies obey this policy.

Credentials and reference URLs

Impact omits signed or authenticated URLs in recognized media and resource reference fields on its outgoing copies, including known native JSON mirrors, even under full content capture. Captured messages and typed payloads also replace recognized authenticated HTTP and relative URLs in prose while retaining surrounding text. Ordinary reference URLs follow the integration’s own reference rules. The Google Models supplement drops the native transport envelope from its capture, including when an Interaction returns the native response. Known OpenAI and Anthropic MCP credentials are excluded as well. Apply your own redaction to arbitrary application attributes and customer-owned OpenTelemetry destinations.

Stream output and usage

Capture observes semantic stream output only as the native owner and caller expose it. It does not read ahead, call a model again or turn partial consumption into a completed response. Completion, caller stop, provider error, cancellation, disconnect and SDK shutdown remain distinct. Terminal identity and usage are known only when the native path exposes them. Dropping an unclosed stream cannot promise terminal output or usage. When Responses supplies cache-read input or reasoning output token counts, capture preserves them as subsets of the inclusive input/output totals. Missing counts remain unknown, and those subsets must not be added to the totals. Selected Product content includes semantic input, output, system instructions, tool definitions, model, response identity, finish reason and usage. It does not promise a raw provider request or response envelope, unknown vendor extensions, or replay of every consumed chunk. Unknown fields emitted by an upstream owner or explicit application capture pass through the telemetry pipeline. See the integration catalogue for other available fields. The SDK reports overflow and serialization loss.

OpenAI structured output and helpers

The OpenAI supplement snapshots stored request schema data at call time in impact.openai.response.schema and records the supported request spelling in impact.openai.response.schema_source as text.format.schema or response_format.json_schema.schema. Request accessors and opaque proxies remain native-client owned. Their values can be unavailable to capture. Native APIPromise shape and helper behavior remain intact, and response parsing remains lazy. An asResponse() body remains untouched and its output is unavailable.raw-response. An unobserved call is settled as unavailable only during bounded SDK shutdown. Early close before a provider terminal signal retains consumed semantic output within a 1 MiB/2,048-part supplement bound and records a partial outcome. A consumed terminal signal preserves its actual outcome. Reaching the bound records partial.local-limit. Standard Responses create streams observe the shared native source once through direct iteration, tee() and native readable conversion. Capture snapshots each observed item before it reaches the application, so later mutation cannot rewrite the recorded output. Shutdown retains the already observed prefix without reading or closing the native stream. Protect-wrapped and resumed retrieve tee paths remain unassessed. This does not retain native chunk replay.

Trace and log content

Semantic input and output logs use the selected logger route. Inspect trace and log delivery independently. The Google supplement emits a typed body for content with own __proto__ or constructor map keys on customer loggers. The Impact destination restores those values to structured attributes. Other Google content keeps its attribute carrier.

Record source locations

Automatic source locations require a configured root and source basis:
Capture observes up to 32 stack frames at the application boundary and emits a safe relative file, available function and line. It never scans, reads or uploads source, and never emits the absolute root. Node native callsites are used when available. The bounded fallback refuses custom stack hooks. Missing frames/mappings are reported as unavailable. Explicit sourcePath/sourceFunction/sourceLine context overrides automatic location coherently. Omit the setting when no trustworthy source basis exists, or use enabled: false to disable automatic capture.