Module catalogue 

The transcribe.cpp catalogue: the per-variant feed the runtime publishes.

Each variant is one JSON document (catalog/<variant>.json) carrying the model’s identity, its downloads, and the measured accuracy and speed rows. Parsing is tolerant by design: upstream grows the schema, so unknown fields are ignored and the optional blocks default, while a document from another schema is refused by name rather than half-read.

Re-exports

use std::io::Read as _; 🔒

use std::path::Path; 🔒

use std::time::Duration; 🔒

use std::time::SystemTime; 🔒

use serde::Deserialize; 🔒

use serde::Deserialize; 🔒

use serde::Serialize; 🔒

use serde::Serialize; 🔒

use crate::Error; 🔒

Structs

AccuracyRow

One measured accuracy row.

Capabilities

What a variant can do, as far as the feed verifies it.

Catalogue

The feed as one fetch found it.

CatalogueEntry

One variant’s catalogue entry.

CatalogueFile 🔒

CatalogueSource

Where the feed is enumerated, fetched and versioned.

Download

One downloadable quantisation of the variant.

HeadlineBenchmark

The benchmark a variant’s headline figure comes from.

License

The licence block, in the two forms the feed carries.

SpeedRow

One measured speed row. Many fields cross that nothing reads; the wall-clock realtime factor is the one the page compares on.

Support

One capability’s support flag.

Enums

EntryKind

What one catalogue entry is FOR: the slot it can fill. Exhaustive on purpose - a consumer that ignores it does not compile, so the two feeds’ entries can never be mixed up by an assumption.

Constants

CATALOGUE_CACHE_FILE 🔒

The record of the last fetched feed under read_catalogue_cache’s directory.

CATALOGUE_CACHE_VERSION 🔒

The record’s own layout version. A file written by a layout this build does not know is ignored rather than misread.

DOWNLOADS_OPEN 🔒

The markers the feed wraps its machine-readable regions in.

FETCH_WORKERS 🔒

How many entry documents are fetched at once. The feed is roughly 75 small files, so this is what turns a ten-second serial walk into one that finishes while the page is still drawing its first frame.

MAX_RESPONSE_BYTES 🔒

Bytes any single catalogue response may carry before the fetch refuses to buffer more. The feed’s documents are about 13 KiB; the cap is what keeps a broken mirror from growing this process.

PREFERRED_DOWNLOADS

The quantisations this runtime runs, best first: the order a download is picked in when a source offers several, and the set a cleanup candidate must offer at least one of. One ladder, so a row the feed admits and the download that row resolves to can never disagree.

REFERENCE_QUANTS 🔒

The quants a headline accuracy figure is read at, most preferred first: the reference quants the catalogue measures its headline rows at, so a comparison across variants reads the same axis.

SCHEMA

The schema string every entry this build reads carries.

TABLE_CLOSE 🔒

Functions

doc_links

The download links one variant’s doc carries, by file name.

download_links

The download links a variant’s own documents carry, by file name.

feed_client 🔒

The client every feed request goes through: the timeouts, and the User-Agent api.github.com refuses a request without.

fetch_catalogue

Fetch the whole catalogue. Blocking, like everything else here.

fetch_chunk 🔒

Fetch and parse one slice of the listing; a document that cannot be read is counted, not returned.

fetch_doc

Fetch one variant’s per-model doc, where its own download links live.

file_links

Every .gguf link in a whole document, by file name.

get_bounded_text 🔒

GET one response, bounded: a status that is not success is its own error, and a body past cap is refused rather than buffered.

gguf_links 🔒

link_targets 🔒

Every markdown link target on one line: the text between ]( and ).

listed_files 🔒

The entry documents a listing names: JSON files, never the underscore-prefixed schema and profile records beside them.

parse_entry

Parse one catalogue document, refusing anything that is not an entry of the schema this build reads.

read_catalogue_cache

Read the last fetched catalogue under dir, when there is one this build wrote.

release_tag 🔒

The tag of a GitHub release document.

rfc3339_now 🔒

Now, in RFC 3339 UTC. A clock before the epoch formats as the epoch rather than failing: a wrong label beats no feed.

write_catalogue_cache

Write the feed under dir, so a page reads the last-known rows without the network.