Module embedding 

One vocabulary and one set of checks for both sides of an embedding exchange.

Structs

Embed

Error

Failed

Input

Limits

Model

Ready

Start

Vectors

Window

EMB-E2: one window of a text: its bytes [start, end), on char and token boundaries, and their vector.

Windows

EMB-E2: how a text longer than the model reads is embedded whole.

Enums

Distance

FailedCode

ItemCode

Kind

Message

Outcome

Tag

Constants

FAILED_MESSAGE_CHARS

MEMORY_LIMIT_EXIT

MINIMUM_FRAME_BYTES

PROTOCOL

PROTOCOL_WINDOWS

EMB-E2: the protocol a Start names when it asks for windows. A worker answers Ready with the protocol its Start named.

Functions

check_embed

Check a batch has a bounded item count and distinct valid identifiers.

check_model_limits

Check that token and frame limits fit the model and its largest answer.

check_model_limits_windowed

check_model_limits, and with windows (EMB-E2) a frame that carries the largest windowed answer and an overlap below the model’s content tokens, so every window moves forward through its text.

check_ready

Ready names protocol 1 or, for a Start that asked for windows, 2; the service checks it names the protocol its Start did.

check_start

Validate the protocol, model path and positive exchange limits.

check_vectors

Check answer count, order, vectors and refusal outcomes against the request.

check_vectors_windowed

check_vectors for a worker started with windows (EMB-E2): an item of a document Embed may also answer its windows, or windows_over_cap; a question is never windowed.

decode_message

Decode a tagged JSON object and validate its exchange envelope.

decode_vector

Decode lowercase little-endian hexadecimal into finite float members.

encode_vector

Encode finite float members as lowercase little-endian hexadecimal.

input_refusal

read_message

Read and decode one bounded frame; clean EOF returns no message.

read_start

Read the first exchange with its fixed bound before allocating the payload.

vectors_message_size

Upper bound for tag plus compact JSON, not the usual message size. Includes maximally escaped ids and the widest whole-number token counts.

windows_message_size

EMB-E2: vectors_message_size when each item may answer up to max_windows windows, each its bounds, its vector and its token count.

write_message

Validate an envelope and write its tagged JSON in one flushed frame.