telemetry-dev-google-genai

Google GenAI (Gemini) SDK instrumentation for telemetry.dev. It wraps the official google-genai Python SDK and emits telemetry.dev generation and embedding spans through telemetry-dev.

Install

pip install telemetry-dev-google-genai

Initialize the core SDK first:

import telemetry_dev

telemetry_dev.init(
    api_key="td_live_...",
    base_url="http://localhost:4318",
    service_name="my-service",
)

Per-client wrapping

from google import genai
from telemetry_dev_google_genai import wrap_google_genai

client = wrap_google_genai(genai.Client(api_key="..."))

client.models.generate_content(
    model="gemini-2.5-flash",
    contents=[{"role": "user", "parts": [{"text": "Tell me a joke about OpenTelemetry"}]}],
)

Use this when you want explicit control over which clients are instrumented. wrap_google_genai patches both sync client.models and async client.aio.models.

Global instrumentation

from google import genai
from telemetry_dev_google_genai import instrument_google_genai, uninstrument_google_genai

instrument_google_genai()
client = genai.Client(api_key="...")

try:
    client.models.generate_content(
        model="gemini-2.5-flash",
        contents="Tell me a joke about OpenTelemetry",
    )
finally:
    uninstrument_google_genai()

Use this as the app-wide one-liner at startup when all GenAI clients should be instrumented.

Instrumented surfaces

Sync and async variants are covered:

• client.models.generate_content(...) / client.aio.models.generate_content(...)
• client.models.generate_content_stream(...) / client.aio.models.generate_content_stream(...)
• client.models.embed_content(...) / client.aio.models.embed_content(...)
• generate_images(...), edit_image(...), upscale_image(...), and generate_videos(...) on sync and async models

client.chats.create(...).send_message(...) and send_message_stream(...) are covered automatically because they call the wrapped models methods.

When generate_videos(...) returns a nonterminal operation, its span completes only after operations.get(...) on the same wrapped client observes that operation in a terminal state. An operation polled by another client or process cannot complete the submission span. If the wrapped client closes, instrumentation is removed, or the bounded operation tracker evicts an uncompleted operation, the submission span ends with an error instead of remaining open.

What gets captured

The integration maps native Gemini request/response shapes directly. It never mutates caller requests.

Tool definitions (gen_ai.tool.definitions, including tool descriptions and parameter schemas) are request attributes, not captured content: capture_input, the mask, and max_attribute_length=0 do not remove them. They are cut only by the OpenTelemetry attribute limit (65,536 characters when max_attribute_length is 0).

Streaming

Gemini streams already include cumulative usage_metadata on chunks, so no request injection is needed. Stream spans record time-to-first-chunk, aggregate text parts (merging consecutive text with the same thought flag), last-seen usage/finish reasons, and end once when the stream completes, errors, or is closed.

Automatic function calling (AFC)

When Python callables are passed in tools, the SDK may run an internal AFC loop across multiple transport calls. The integration emits one span for the public generate_content call and, when present, sets span input to automatic_function_calling_history.

Embeddings

Embedding calls emit gen_ai.operation.name = "embeddings", request model/input, embedding count/dimension attributes, optional token usage, and optional billable character counts. Embedding vectors are not captured as output.

Provider values

• Gemini Developer API clients record provider gcp.gemini.
• Vertex AI clients (vertexai=True) record provider gcp.vertex_ai.

Fail-open guarantee

Mapping code is defensive; wrapped calls return the SDK response unchanged and re-raise exceptions untouched. Telemetry bugs never break callers.

Limitations

• Not instrumented: count_tokens, compute_tokens, live, caches, files, tunings, batches.
• Media operations capture text prompts, model/provider attribution, non-binary media references such as URIs and MIME types, and gen_ai.operation.name=generate_content. Image, video, and reference-image bytes are never captured. Image and video operations emit the corresponding gen_ai.output.type.
• Deliberately not captured: logprobs, citation metadata, create_time, sdk_http_response.
• Unconsumed streams end their spans only when the stream is exhausted, errors, or is closed.