LangSmith

Captures hierarchical execution traces across LLM calls, chain steps, and agent actions by instrumenting LangChain runtime via SDK hooks and context propagation. Traces include token counts, latencies, inputs/outputs, and error states, visualized as interactive DAGs showing call dependencies and performance bottlenecks. Uses span-based tracing architecture similar to OpenTelemetry but optimized for LLM-specific metadata (model names, temperature, token usage).

Centralized registry for storing, versioning, and deploying LLM prompts with git-like commit history, branching, and rollback capabilities. Prompts are stored as immutable versions linked to evaluation results and production deployments. Supports templating with Jinja2 or Handlebars for dynamic variable injection, and integrates with LangChain's LLMChain to pull prompts at runtime via semantic versioning (e.g., 'my-prompt@latest' or 'my-prompt@v2.3').

Monitors trace metrics (latency, error rate, token usage, cost) in real-time and triggers alerts when metrics exceed thresholds or deviate from baseline patterns. Uses statistical anomaly detection (z-score, moving average) to identify unusual behavior without manual threshold configuration. Supports multiple notification channels (email, Slack, webhooks) and integrates with incident management platforms.

Exposes REST and GraphQL APIs for querying traces, running evaluations, managing datasets, and accessing evaluation results programmatically. Enables building custom dashboards, integrating with external analysis tools, or automating evaluation workflows. APIs support filtering, pagination, and bulk operations. Authentication via API keys with role-based access control.

Manages labeled datasets (inputs, expected outputs, metadata) and runs evaluation jobs that execute chains against dataset examples, computing both built-in metrics (exact match, token overlap, semantic similarity via embeddings) and custom Python-defined metrics. Evaluation results are aggregated into scorecards showing pass rates, latency distributions, and cost breakdowns per model or prompt version. Supports batch evaluation with configurable concurrency and retry logic.

Provides a web UI for human annotators to review LLM outputs from production traces, assign labels (correct/incorrect, quality ratings, category tags), and add free-form feedback. Annotations are stored as structured records linked to the original trace and can be exported as labeled datasets for fine-tuning or retraining evaluation models. Supports collaborative workflows with role-based access (viewer, annotator, admin) and bulk operations for labeling multiple examples.

Automatically extracts and aggregates token counts and API costs from LLM calls across multiple providers (OpenAI, Anthropic, Cohere, Azure, local models) by parsing model names and pricing tables. Provides dashboards showing cost per trace, per user, per prompt version, and per model, with drill-down capabilities to identify expensive chains. Supports custom pricing rules for self-hosted or fine-tuned models. Costs are calculated in real-time during trace collection and stored with each span.

Groups traces by user ID, session ID, or custom tags to enable conversation-level and user-level analysis. Provides session timelines showing all traces for a user in chronological order, with filtering by date range, model, or trace status. Supports session-level metrics (total cost, total tokens, conversation length) and enables bulk operations (e.g., export all traces for a user, delete traces for a user). Session data is indexed for fast retrieval and supports multi-tenant isolation.

Provides built-in benchmarking workflows to compare models, prompt versions, or configurations on the same dataset with statistical significance testing. Generates comparison reports showing latency distributions, token efficiency, cost per output, and custom metric scores with confidence intervals. Supports A/B testing with automatic traffic splitting and statistical power analysis to determine required sample size for significance.

Provides language-specific SDKs (Python, JavaScript/TypeScript) that automatically instrument LangChain chains and agents with minimal code changes. Uses context variables and decorators to capture execution context without modifying chain logic. Supports both synchronous and asynchronous execution, with automatic error handling and retry logic. Traces are batched and sent asynchronously to avoid blocking application execution.

Abstracts LLM provider differences (OpenAI, Anthropic, Cohere, Azure, local models) through a unified tracing interface that captures provider-specific metadata (model name, temperature, top_p, token limits) consistently. Automatically maps provider-specific response formats to a standard trace schema, enabling cross-provider comparison and cost tracking. Supports streaming responses with token-by-token tracing.

Enables feedback collection from production traces (thumbs up/down, ratings, free-form comments) and automatically exports labeled examples to create fine-tuning datasets. Integrates with evaluation runs to track how model performance changes over time as new feedback is collected. Supports feedback aggregation by user, model, or prompt version to identify improvement opportunities.