@tavily/ai-sdk

Executes semantic web searches that understand query intent and return contextually relevant results with source attribution. The SDK wraps Tavily's search API to provide structured search results including snippets, URLs, and relevance scoring, enabling AI agents to retrieve current information beyond training data cutoffs. Results are formatted for direct consumption by LLM context windows with automatic deduplication and ranking.

Extracts structured, cleaned content from web pages by parsing HTML/DOM and removing boilerplate (navigation, ads, footers) to isolate main content. The extraction engine uses heuristic-based content detection combined with semantic analysis to identify article bodies, metadata, and structured data. Output is formatted as clean markdown or structured JSON suitable for LLM ingestion without noise.

Crawls websites by following links up to a specified depth, extracting content from each page while respecting robots.txt and rate limits. The crawler maintains a visited URL set to avoid cycles, extracts links from each page, and recursively processes them with configurable depth and breadth constraints. Results are aggregated into a structured format suitable for knowledge base construction or site mapping.

Analyzes a website's link structure to generate a navigational map showing page hierarchy, internal link density, and site topology. The mapper crawls the site, extracts all internal links, and builds a graph representation that can be visualized or used to understand site organization. Output includes page relationships, depth levels, and link counts useful for navigation-aware RAG or site analysis.

Provides Tavily tools as composable functions compatible with Vercel AI SDK's tool-calling framework, enabling automatic serialization to OpenAI, Anthropic, and other LLM function-calling APIs. Tools are defined with JSON schemas that describe parameters and return types, allowing LLMs to invoke search, extraction, and crawling capabilities as part of agent reasoning loops. The SDK handles parameter marshaling, error handling, and result formatting automatically.

Streams search results, extracted content, and crawl findings progressively as they become available, rather than buffering until completion. Uses server-sent events (SSE) or streaming JSON to yield results incrementally, enabling UI updates and progressive rendering while operations complete. Particularly useful for crawls and extractions that may take seconds to complete.

Provides structured error handling for network failures, rate limits, timeouts, and invalid inputs, with built-in fallback strategies such as retrying with exponential backoff or degrading to cached results. Errors are typed and include actionable messages for debugging, and the SDK supports custom error handlers for application-specific recovery logic.

Handles Tavily API key initialization, validation, and secure storage patterns compatible with environment variables and secret management systems. The SDK validates keys at initialization time and provides clear error messages for missing or invalid credentials. Supports multiple authentication patterns including direct key injection, environment variable loading, and integration with Vercel's secrets management.