@lacspace/tokenizer

A tiny, zero-dependency, provider-agnostic LLM chat client over fetch — one chat()/stream() API for OpenAI, Anthropic, Google Gemini and any OpenAI-compatible endpoint (Groq, Together, OpenRouter, Ollama), plus typed message builders, retry/backoff + timeout, usage/cost accounting and JSON-output parsing. Bring your own key, no SDK, isomorphic.

A tiny, typed prompt-engineering toolkit — build, compose and render LLM prompts and chat messages with real type safety (variable names inferred from the template). Now with a versioned prompt registry, prompt-injection guards, JSON/enum output-format instructions and pure token-budget trimming. Zero-dependency, isomorphic.

Extract and repair JSON from messy LLM output — strip code fences and prose, pull out every embedded JSON value, repair with typed diagnostics, and fix trailing commas, single quotes, unquoted keys, Python literals and truncated objects. Zero-dependency, isomorphic, typed.

A tiny text splitter for RAG and long-context prompts — recursive character, Markdown, code, sentence and paragraph chunking with overlap and a pluggable token-aware lengthFn. Now batch-chunks whole corpora (per-chunk docId + metadata), merges tiny fragments and ships a dependency-free approximate token counter. Zero-dependency, isomorphic.

A zero-dependency, isomorphic parser for Server-Sent Events, NDJSON and streaming LLM responses — a unified async iterator of text and tool-call deltas across OpenAI and Anthropic, plus map/filter/tee transforms, ReadableStream adapters and AbortSignal cancellation. No SDK, keyless.

Define LLM function-calling / tool-use tools once and use them with any provider — generate the OpenAI, Anthropic or Google tool spec, parse tool calls out of any response, validate/coerce the model's arguments (with an opt-in strict validator), dispatch to your handler and format results back into messages. Zero-dependency, keyless.