Observability Correlation

AgentBridge keeps LangSmith and Langfuse as optional integrations, but uses one metadata contract for LangChain and LangGraph runs. observability_metadata() merges application metadata with session, user, project, and trace identifiers without storing provider secrets in AgentSpec.

from agentbridge.observability import observability_metadata

metadata = observability_metadata(
    {"observability": {
        "langsmith": {"enabled": True, "project_name": "refunds"},
        "langfuse": {"enabled": True, "user_id": "customer-123"},
    }},
    session_id="session-1",
)

The same metadata is forwarded into LangChain and LangGraph runtime configuration. LangSmith tracing context and Langfuse callbacks remain lazy and optional; credentialed API smoke tests are separate from the credential-free correlation contract.

Langfuse Telemetry API

The LangChain plugin exposes the current OTLP/HTTP path and retains the legacy JSON ingestion helper only as a compatibility escape hatch:

from agentbridge_langchain.langfuse_api import LangfuseAPIClient

client = LangfuseAPIClient()
client.ingest_otlp({"resourceSpans": []})
client.list_observations(query={"limit": 100})
client.list_scores_v3(query={"dataType": "NUMERIC"})
list(client.iter_observations(query={"traceId": "trace-123", "limit": 100}))
list(client.iter_scores_v3(query={"dataType": "NUMERIC", "limit": 100}))
client.query_metrics(query={"view": "traces"})
client.list_experiments(query={"fields": "core,scores"})
client.list_experiment_items(query={"experimentId": "exp-123", "fields": "io,scores"})
client.list_datasets(query={"limit": 50})
client.list_dataset_items(query={"datasetName": "refunds"})

ingest_otlp() also accepts protobuf bytes with content_type="application/x-protobuf" and sends x-langfuse-ingestion-version: 4 by default. The iterators follow meta.cursor across pages while preserving the caller's filters. The generic request methods remain available for new Langfuse endpoints without waiting for a core release.

Dataset lifecycle helpers use the current v2 dataset routes and the versioned dataset-item API: create_dataset, list_datasets, get_dataset, delete_dataset, and item create/upsert/list/get/ delete operations. Legacy trace reads remain available only as a compatibility escape hatch; new trace extraction should query Observations v2 with a traceId filter. Trace deletion helpers are also available for cleanup workflows.

Langfuse Prompts

Prompt versions and deployment labels can be fetched without coupling AgentBridge core to the Langfuse SDK:

from agentbridge_langchain.langfuse_prompts import fetch_prompt

prompt = fetch_prompt(client, "refund-policy", label="production")
text = prompt.compile(customer_tier="gold")
chat_messages = prompt.get_langchain_prompt()

The wrapper preserves the raw response, supports text and chat prompts, and leaves unresolved Langfuse variables as LangChain-style {name} placeholders.