AI SDK Patterns

Create dynamic tools with runtime-generated schemas using dynamicTool. Demonstrates unit conversion tools where schema adapts based on conversion type (temperature, length, weight, currency) and supports dynamic model selection.

Scrape websites using Cheerio's fast HTML parser. Includes URL scraping, website crawling, and content extraction without external API costs.

Resolve messy card statement descriptors to the merchants behind them. Code blocks and vetoes, Jev answers one boolean per pair, and unsure matches wait for review.

Discover themes in comments with no labels up front. Embeddings pick likely pairs, Jev answers one boolean per pair, and code does the grouping.

Run a rubric eval suite with jev-evals: one Jev call per case answers every rubric, and a baseline diff catches fractional regressions.

A FAQ whose last row is a live ask field. Jev opens a canned answer when it matches, or a short grounded answer becomes another row.

Long-term agent memory with three Jev gates. Keep durable facts, recall what fits a token budget, and evict what went stale. Nothing is rewritten.

Prune an agent's tool list before the model sees it. Allow and deny rules run first; Jev asks one relevance boolean per remaining tool in a single call.

Exa finds live evidence, Jev answers the same question with and without it, and leave-one-out weights show what each source changed.

See if stock covers an open order. The model writes one program that calls both lookups and subtracts.

One command at a time. It passes the gate only when a Jev boolean clears 0.90.

Paste a URL. Firecrawl fetches the page, then Jev scores whether the answer block is worth citing.

Three prepared replies sit on the table. A Jev choice lifts the winner only when its probability clears 0.70.

Score a pull request description with three atomic Jev score questions, then combine the answers with weights in code.

Four trays are on the table. A message flies to a team only when confidence and the selected probability both clear their floors.

Tag a closed inbox with one Jev boolean per tag. A stamp lands only when its probability clears a floor.

Look up facts one step at a time. Stop when Jev is at least 90% sure the notes answer the task, or after 3 lookups.

Write an answer, then show it only if Jev is at least 90% sure it actually answers the question. If not, rewrite once.

Filter a closed catalog with one Jev boolean per item. Matches rise out of the pile when their probability clears a floor.

Pick one specialist with a Jev choice, then run only that agent when confidence and selected probability both clear their floors.

Filter or sort a closed table with one Jev question per row. A column name never leaves the app.

Caption a curated image, embed the caption, and export a portable image index without a vector database.

Search a portable image index in the browser with caption embeddings — no vector database.

Compare small, large, and paragraph-window chunking on one document with cosine similarity — no vector database.

Load a sample Local RAG index, rank chunks in the browser, and stream a cited answer without a vector database.

Chunk local text, create embeddings, and export a portable RAG index without a vector database.

Save and recall facts with Supermemory tools — addMemory and searchMemories in a simple Remember / Recall playground.

Paste CSV, TSV, or JSON arrays; detect delimiters with Papa Parse; validate rectangular rows; preview in a table; export JSON, CSV, TSV, or Markdown tables in the browser.

Parse markdown into H1 sections with unified and remark, then export normalized content as JSON, CSV, Markdown, HTML, PDF, or plain text in the browser.

Load Excel .xlsx or .xls workbooks in the browser with SheetJS, pick a sheet, preview rows, and download CSV—with optional UTF-8 BOM and multi-sheet export.

Search with Exa AI SDK's webSearch tool and Anthropic Claude Sonnet 4.5. Uses @exalabs/ai-sdk for seamless integration.

Pass arbitrary context from generateText/streamText to tool execution via experimental_context. Access context in tool execute functions for dynamic behavior.

Explore tool input lifecycle hooks: onInputStart, onInputDelta, and onInputAvailable. See how to track tool input generation in real-time during streaming.

Return preliminary tool results using async iterables. Stream status updates during tool execution before the final result is available.

Repair invalid tool calls using experimental_repairToolCall. Fix schema validation errors without requiring additional steps that pollute message history.

Search the web with Claude AI. Includes result filtering, multiple search types, and language support.

Search with Exa's semantic API and real-time crawling. Includes live content extraction and relevance scoring.

Crawl and extract content using Firecrawl's API. Includes URL processing, markdown conversion, and structured data extraction.

Process PDFs with AI analysis. Extract text, generate summaries, and identify key insights.

Search the web using OpenAI's tool with real-time results. Includes source verification, citation tracking, and contextual answers.

Scrape websites using Jina AI with JavaScript rendering and proxy rotation. Handles rate limiting and bot detection.

Extract content from websites using Cheerio. Includes parsing, link discovery, and rate limiting.

Convert any webpage to markdown using Cloudflare's markdown.new. Clean, structured output for AI pipelines and content processing.

Scrape websites using Jina AI's reader service with JavaScript rendering. Includes URL scraping, website crawling, and content extraction with AI-powered parsing.

Scrape websites using Cloudflare's markdown.new service. Extract clean markdown content from any URL. No API key required.

Search the web using Exa's semantic API and real-time crawling. Includes URL search, content extraction, Wikipedia, Reddit, and news search tools.

Search the web using Firecrawl's API and real-time crawling. Includes URL search, content extraction, Wikipedia, Reddit, and news search tools.