tokvera 0.2.3

Lightweight multi-provider (OpenAI, Anthropic, Gemini) analytics tracking wrapper

tokvera

tokvera is a lightweight Python SDK that wraps OpenAI, Anthropic, and Gemini clients and emits usage analytics in a fire-and-forget way.

What's New in v0.2.2

• Added Trace Context v1 tags.
• New optional tags: trace_id, run_id, conversation_id, span_id, parent_span_id, step_name.
• Added Evaluation Signals v1 fields: outcome, retry_reason, fallback_reason, quality_label, feedback_score.
• Added FastAPI middleware integration helpers.
• Added LangChain callback integration helpers.
• Added LlamaIndex callback integration helpers.
• Auto-generates trace_id and span_id when you do not provide them.

Installation

pip install tokvera

For development:

pip install -e .[dev]

Environment Variable Setup

Set your ingestion endpoint:

# Linux/macOS
export TOKVERA_INGEST_URL="https://api.tokvera.com/v1/events"

# Windows PowerShell
$env:TOKVERA_INGEST_URL = "https://api.tokvera.com/v1/events"

If TOKVERA_INGEST_URL is not set, analytics are skipped automatically.

Trace Context v1

Use trace tags to reconstruct request chains without sending prompt payloads.

Recommended semantics:

• trace_id: one end-to-end workflow/request.
• run_id: one execution run of an agent/workflow.
• conversation_id: one user conversation/session.
• span_id: one model call.
• parent_span_id: parent model call when nested.
• step_name: readable stage label (retrieve_context, draft_reply, quality_retry).

Example:

client = track_openai(
    openai_client,
    api_key="tokvera_project_key",
    feature="support_bot",
    tenant_id="acme",
    trace_id="trace_req_20260304_001",
    run_id="run_agent_20260304_001",
    conversation_id="conv_9832",
    span_id="span_root_1",
    parent_span_id=None,
    step_name="draft_reply",
)

FastAPI Middleware Integration

Use middleware to create request-level trace context and pass it into SDK calls.

from fastapi import FastAPI, Request
from openai import OpenAI
from tokvera import (
    create_fastapi_tracking_middleware,
    get_fastapi_track_kwargs,
    track_openai,
)

app = FastAPI()
openai_client = OpenAI(api_key="sk-...")

middleware = create_fastapi_tracking_middleware(
    defaults={"feature": "support_bot", "environment": "production"},
    context_resolver=lambda request: {"tenant_id": request.headers.get("x-tenant-id")},
)

@app.middleware("http")
async def tokvera_context(request: Request, call_next):
    return await middleware(request, call_next)

@app.post("/reply")
async def reply():
    tracked = track_openai(
        openai_client,
        api_key="tokvera_project_key",
        **get_fastapi_track_kwargs(step_name="draft_reply"),
    )
    return tracked.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": "Hello"}],
    )

LangChain Callback Integration

Use a callback handler to emit Tokvera events from LangChain LLM runs.

from langchain_openai import ChatOpenAI
from tokvera import create_langchain_callback_handler

callback = create_langchain_callback_handler(
    api_key="tokvera_project_key",
    feature="agent_support",
    tenant_id="acme",
    environment="production",
)

model = ChatOpenAI(
    model="gpt-4o-mini",
    callbacks=[callback],
)

result = model.invoke("Hello")

LlamaIndex Callback Integration

Use a callback handler to emit Tokvera events from LlamaIndex workflows.

from llama_index.core.callbacks import CallbackManager
from tokvera import create_llamaindex_callback_handler

tokvera_handler = create_llamaindex_callback_handler(
    api_key="tokvera_project_key",
    feature="agent_support",
    tenant_id="acme",
    environment="production",
)

callback_manager = CallbackManager([tokvera_handler])

Quick Start

OpenAI

from openai import OpenAI
from tokvera import track_openai

openai_client = OpenAI(api_key="sk-...")

client = track_openai(
    openai_client,
    api_key="tokvera_project_key",
    feature="support_bot",
    tenant_id="acme",
    trace_id="trace_support_001",
    run_id="run_support_001",
    conversation_id="conv_42",
    step_name="draft_reply",
    outcome="success",
    quality_label="good",
    feedback_score=5,
    plan="pro",
    environment="production",
    template_id="support_v3",
    capture_content=False,
)

response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Hello"}],
)

Anthropic

from anthropic import Anthropic
from tokvera import track_anthropic

anthropic_client = Anthropic(api_key="sk-ant-...")

client = track_anthropic(
    anthropic_client,
    api_key="tokvera_project_key",
    feature="support_bot",
    tenant_id="acme",
    environment="production",
)

client.messages.create(
    model="claude-3-5-sonnet-latest",
    max_tokens=256,
    messages=[{"role": "user", "content": "Hello"}],
)

Gemini

from google import genai
from tokvera import track_gemini

gemini_client = genai.Client(api_key="AIza...")

client = track_gemini(
    gemini_client,
    api_key="tokvera_project_key",
    feature="assistant",
    tenant_id="acme",
    environment="production",
)

client.models.generate_content(
    model="gemini-2.0-flash",
    contents="Hello",
)

Event Schema

Canonical specification: tokvera-api/docs/EVENT_SCHEMA.md

Events include:

• schema_version: 2026-02-16
• event_type: openai.request, anthropic.request, or gemini.request
• provider: openai, anthropic, or gemini
• endpoint: chat.completions.create, responses.create, messages.create, models.generate_content
• status: success or failure
• latency_ms
• model
• usage: prompt_tokens, completion_tokens, total_tokens
• tags: feature, tenant_id, customer_id, attempt_type, plan, environment, template_id, trace_id, run_id, conversation_id, span_id, parent_span_id, step_name
• Evaluation signals (optional): outcome, retry_reason, fallback_reason, quality_label, feedback_score (emitted in tags and top-level evaluation)
• error on failure events

trace_id and span_id are auto-generated per request if not provided.

Privacy

By default, prompt/response content is not sent.

If capture_content=True, content is hashed (SHA-256) before ingestion. Raw content is never sent by this SDK.

Disable Tracking

You can disable tracking by either:

1. Using the original OpenAI client directly (do not wrap it), or
2. Unsetting TOKVERA_INGEST_URL so no events are emitted.