@plurnk/plurnk-mimetypes-application-gguf

application/x-gguf (GGUF local-LLM model) mimetype handler for the plurnk ecosystem. Hand-rolled binary reader, no parser dependency.

install

npm i @plurnk/plurnk-mimetypes-application-gguf

what it does

GGUF (llama.cpp / Ollama) packs a model's weights behind a structured key-value metadata header: architecture, quantization, context length, embedding length, tokenizer, chat template. This handler reads only that header — never the gigabytes of tensor data — so an agent in a model folder can answer "what model is this, what quant, what context window" for free.

• extractRaw(content) — the metadata keys (general.architecture, llama.context_length, …) as field symbols, in file order.
• deepJson(content) — { version, tensorCount, metadata }, a jsonpath target ($.metadata['llama.context_length']).
• toText (regex/glob + embed-source) — the metadata rendered as a key: value table. Large arrays (tokenizer vocabularies) are summarized as <type[N]>, never expanded.
• validate — throws on bad magic or a truncated header; every other channel degrades to empty.

Per-tensor inventory is out of scope for v1 — the metadata is the "what is this" answer. References are not applicable.

license

MIT.