BigQuery Agent Analytics SDK

An open-source Python SDK for analyzing, evaluating, and curating agent traces stored in BigQuery. Built on top of the BigQuery Agent Analytics, it provides a consumption-layer toolkit for agent observability, analysis, evaluation, and advanced capabilities like the Agent Context Graph — extracting decision traces from your agent's context graph — at scale.

Overview

The BigQuery Agent Analytics SDK connects your AI agent telemetry in BigQuery to a rich set of evaluation, observability, and analytics capabilities. It is designed for ML engineers, data scientists, and platform teams who run agents in production and need to understand agent behavior, measure quality, and detect regressions — all through BigQuery SQL or Python.

Key Features

Observability

• Trace reconstruction and DAG visualization
• Per-event-type BigQuery views
• Observability dashboards (Looker Studio, SQL, and BigFrames)

Evaluation

• System metrics (latency, turn count, tool call error rate, token efficiency, time to first token, cost)
• Performance Metrics (correctness, hallucination, sentiment, efficiency, etc)
• Multi-trial system and performance metrics
• Grader composition (weighted, binary, majority strategies)
• Eval suite lifecycle management with graduation and saturation detection
• Static quality validation (ambiguous tasks, class imbalance, suspicious thresholds)

AI/ML Integration

• BigQuery AI.GENERATE, AI.EMBED, AI.CLASSIFY
• Anomaly detection and latency forecasting
• Categorical (Hatteras-style) evaluation via BigFrames

Advanced Analytics

• Agent Context Graph — extract decision traces from your agent's context graph: the requests an agent handled, the options it weighed, and the outcomes it committed, materialized into a queryable BigQuery property graph (GQL traversal, scheduled refresh via bqaa context-graph)
• Long-horizon cross-session memory
• Multi-stage agent insights pipeline
• Drift detection for golden vs production question distributions

CLI (bq-agent-sdk)

• 12+ commands for diagnostics, evaluation, and CI/CD integration

Deployment Surfaces

• Remote Function (BigQuery SQL via Cloud Run)
• Python UDF scoring kernels
• Streaming evaluation (Cloud Scheduler + Cloud Run)
• Continuous query templates

Usage Telemetry

• Every job the SDK submits is labeled (sdk, sdk_version, sdk_surface, sdk_feature, and sdk_ai_function where relevant) so operators can attribute spend, latency, and adoption directly from INFORMATION_SCHEMA.JOBS_BY_PROJECT. No extra telemetry pipeline is required. See docs/sdk_usage_tracking.md for the label schema and ready-to-run tracking queries.

How this fits together

Two kinds of components share one agent_events table:

• Producers write it. The ADK BigQueryAgentAnalyticsPlugin and the LangChain / LangGraph BigQueryCallbackHandler emit the same schema and can write to the same table side by side.
• This SDK consumes it. Everything here is read-only over that table: trace reconstruction, evaluation, insights, the context graph.

Prerequisites

• Python 3.10+
• A Google Cloud project with BigQuery enabled
• Agent traces already written to BigQuery by one of the producers above

Installation

pip install bigquery-agent-analytics

With optional LLM judge support:

pip install bigquery-agent-analytics[llm]

With BigFrames support:

pip install bigquery-agent-analytics[bigframes]

With LangSmith export support:

pip install bigquery-agent-analytics[langsmith]

Quick Start

from bigquery_agent_analytics import Client, TraceFilter

client = Client(
  project_id="my-project",
  dataset_id="analytics",
  table_id="agent_events",  # the table your producer writes (this is the default)
  location="US",  # your dataset's location; omit to use the BigQuery client default
)

traces = client.list_traces(TraceFilter(limit=1))
if traces:
  traces[0].render()
else:
  print("No traces found. Check that agent events have been ingested.")

If the table or location does not match what your producer wrote, reads raise EventsTableNotFoundError, whose message names both and says what to pass. It is a google.api_core.exceptions.NotFound, so existing handlers still catch it.

Export traces to LangSmith

Export the standard ADK agent_events schema with Application Default Credentials and LangSmith's standard environment variables:

export LANGSMITH_API_KEY=lsv2_...
export LANGSMITH_PROJECT=agent-production

bq-agent-sdk export langsmith \
  --source=my-project.analytics.agent_events \
  --since=2026-08-01T00:00:00Z

The CLI intentionally accepts the API key only through LANGSMITH_API_KEY, keeping it out of shell history and process arguments.

The exporter reconstructs span parents and derives stable LangSmith UUIDs. It treats created runs as immutable: replaying an overlapping window is an idempotent no-op for existing run IDs while still creating previously unseen IDs. Run IDs derive from the source identity and not the destination, so exporting to a fresh LangSmith project reuses the same IDs and creates nothing. To correct already-exported data, use a deliberately versioned --source-id. For scheduled syncs, use --incremental with --watermark-file=state.json. The JSON summary bounds row-level diagnostics with --max-dropped-rows and reports the number omitted as dropped_rows_truncated.

LangSmith Cloud rejects runs whose start_time is more than 24 hours from now, so an export covers recent traces rather than aged trace history. Run --incremental on a schedule frequent enough to stay inside that window. See SDK.md for the exact error and its effect on bounded backfills.

Custom schemas use a YAML mapping from LangSmith fields to source column or nested paths. Unmapped columns remain in extra.metadata; payload values are opaque and are never classified by event type. Optional fields omitted from a custom mapping remain unmapped rather than inheriting ADK column names:

fields:
  run_id: event_key
  trace_id: trace.key
  parent_run_id: parent_key
  name: kind
  start_time: occurred_at
  inputs: payload

bq-agent-sdk export langsmith \
  --source='SELECT * FROM `my-project.custom.events`' \
  --mapping=mapping.yaml --source-id=custom-events-v1

See SDK.md for the Python API, incremental watermark contract, filtering, and operational controls.

For session reads, session_id is a reusable conversation identifier rather than a unique trace key. client.get_session_trace() resolves user, root agent, experiment, and labels; if more than one candidate remains it raises AmbiguousSessionError carrying structured candidates for an exact client.get_trace_by_selector() retry. The same contract is used by GQL, trajectory evaluation, the CLI, the Remote Function, and reports. See Identity-safe session resolution.

Categorical evaluation can bind trusted per-trace judge context (for example, a golden expected answer) to the same exact selector:

from bigquery_agent_analytics import ResolvedTraceSelector, TraceFilter

filters = TraceFilter(limit=100)
traces = client.list_traces(filters)
context = {
    ResolvedTraceSelector(trace.identity, trace.scope): expected_answer(trace)
    for trace in traces
}
report = client.evaluate_categorical(
    config,
    filters=filters,
    per_session_context=context,
)

Legacy string keys are accepted only when the transcript-eligible evaluated session_id is unambiguous; eligibility is applied before that ambiguity check, so exact selector keys are recommended whenever session IDs may be reused. Otherwise AmbiguousSessionError fails before any model call. Context is trusted evaluator material, sent as a query parameter/model prompt through AI.GENERATE, retry, and API fallback. It is never interpolated into SQL, logged, persisted, or placed in job labels. Apply the same data-governance policy you use for evaluation prompts.

When persist_results=True, categorical results use an additive, nullable identity/provenance schema; existing historical rows are not backfilled. Deploy or roll back safely in this order: schema, then writer, then views. The latest-results view keeps identities distinct even when they share a session_id. During a legacy/schema straddle, a sole typed identity supersedes matching legacy metric/prompt rows; zero or multiple typed identities leave legacy rows in their separate legacy:<session_id> lane. Trusted judge or golden-answer context — including any model echo — is never persisted; only SDK-owned context provenance is. This U5 migration completes #358's remaining persistence/report gate and unlocks U6/#360.

See SDK.md for the full API walkthrough with code examples for every feature.

Try it: extract decision traces (Agent Context Graph, ~10 minutes)

Deploy a context graph, seed sample agent events, extract the decision traces, and query one in GQL — entirely from your terminal:

export PROJECT_ID="your-project" DATASET="agent_analytics_demo"
gcloud config set project "$PROJECT_ID"
bq --location=US mk --dataset "$PROJECT_ID:$DATASET"

# 1. Deploy the context graph (one-time DDL: tables, then the property graph).
cd examples/context_graph/codelab
envsubst < table_ddl.sql      | bq query --use_legacy_sql=false
envsubst < property_graph.sql | bq query --use_legacy_sql=false

# 2. Seed five sample agent sessions into agent_events.
bqaa seed-events --project-id "$PROJECT_ID" --dataset-id "$DATASET" --sessions 5

# 3. Extract decision traces from the deployed graph
#    (read back via INFORMATION_SCHEMA.PROPERTY_GRAPHS — no SQL file passed).
bqaa context-graph --project-id "$PROJECT_ID" --dataset-id "$DATASET" \
    --graph agent_decisions_graph --lookback-hours 24 --format json

# 4. Query a decision trace: what did the agent weigh, and how did it resolve?
bq query --use_legacy_sql=false "
SELECT * FROM GRAPH_TABLE(
  $DATASET.agent_decisions_graph
  MATCH (req:DecisionRequest)-[eo:evaluatesOption]->(opt:DecisionOption),
        (req)-[ri:resultedIn]->(out:DecisionOutcome)
  COLUMNS (req.request_text AS question, opt.option_label AS considered,
           out.status AS outcome, out.rationale AS rationale))"

Expect "ok": true with 5 sessions materialized, and fifteen GQL rows — three options weighed per request, each with the committed outcome and rationale. The Agent Context Graph codelab is the guided version of these steps (plus backfill and production scheduling), and examples/context_graph/ is the worked example with a runnable ADK agent.

Documentation

Architecture

src/bigquery_agent_analytics/
│
├── Core
│   ├── client.py                  # High-level SDK client
│   ├── trace.py                   # Trace reconstruction & visualization
│   ├── views.py                   # Per-event-type BigQuery view management
│   ├── event_semantics.py         # Canonical event type helpers & predicates
│   ├── serialization.py           # Uniform serialization layer
│   └── formatter.py               # Output formatting (json/text/table)
│
├── Evaluation
│   ├── system_evaluator.py        # SystemEvaluator
│   ├── performance_evaluator.py   # PerformanceEvaluator
│   ├── multi_trial_performance_evaluator.py # MultiTrialPerformanceEvaluator
│   └── aggregate_grader.py        # AggregateGrader
│   ├── eval_suite.py              # Eval suite lifecycle management
│   └── eval_validator.py          # Static validation checks
│
├── AI/ML
│   ├── ai_ml_integration.py       # BigQuery AI/ML capabilities
│   ├── bigframes_evaluator.py     # BigFrames DataFrame evaluator
│   ├── categorical_evaluator.py   # Hatteras categorical evaluation
│   └── categorical_views.py       # Categorical metric views
│
├── Analytics
│   ├── insights.py                # Multi-stage insights pipeline
│   ├── feedback.py                # Drift detection & question distribution
│   └── memory_service.py          # Long-horizon agent memory
│
├── Export
│   └── export/
│       ├── __init__.py             # Stable public export API
│       ├── cli.py                  # bq-agent-sdk export command group
│       └── langsmith.py            # Schema-agnostic LangSmith connector
│
├── Agent Context Graph
│   ├── context_graph.py           # Decision-trace extraction & GQL traversal
│   ├── materialize_window.py      # Scheduled materialization (bqaa context-graph)
│   └── property_graph_spec.py     # Derive the spec from your deployed property graph
│
└── CLI & Deploy
    ├── cli.py                     # CLI entry point (bq-agent-sdk)
    ├── udf_kernels.py             # Python UDF scoring kernels
    └── udf_sql_templates.py       # UDF SQL generation

Related Projects

• Google ADK — Agent Development Kit for building AI agents
• BigQuery — Google Cloud analytics data warehouse
• BigQuery AI Functions — AI.GENERATE, AI.EMBED, AI.CLASSIFY, and more

Development

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Format code
pyink --config pyproject.toml src/ tests/
isort src/ tests/

Contributing

See CONTRIBUTING.md for guidelines.

License

Apache License 2.0 — see LICENSE for details.