Struct LlmEmbedding 

pub struct LlmEmbedding { /* private fields */ }

Implementations

impl LlmEmbedding

pub fn detect_available() -> Result<Self, AppError>

Detects which LLM CLI is available on PATH and returns the matching embedding client.

v1.0.76: PREFERS codex over claude because:

• Claude Code 2.1+ ships a 180k+ token system context (plugins, skills, agents, MCP) that overflows the 200k context window for even trivial embedding prompts and returns “Prompt is too long”. (v1.0.79/S6 mitigates this with an empty CLAUDE_CONFIG_DIR, but codex stays the lighter default.)
• Codex 0.134+ is lightweight (~5k system context) and the StructuredOutput tool reliably returns the requested vectors.

pub fn with_codex() -> Result<Self, AppError>

pub fn with_claude() -> Result<Self, AppError>

pub fn embed_passage(&self, text: &str) -> Result<Vec<f32>, AppError>

Embeds a single passage (chunk of a memory body). Returns an f32 vector of the active dimensionality.

pub fn embed_query(&self, text: &str) -> Result<Vec<f32>, AppError>

Embeds a single query. The LLM uses a different prompt prefix to disambiguate query from passage.

pub async fn embed_batch_async( &self, prefix: &str, batch: &[(usize, String)], ) -> Result<Vec<(usize, Vec<f32>)>, AppError>

G42/S2: embeds a batch of (global_index, text) pairs in ONE LLM call. Returns (global_index, vector) pairs. Async — this is the unit of work scheduled by the bounded fan-out in crate::embedder.

Cancel safety: the future owns its subprocess via kill_on_drop(true), so dropping it (e.g. losing a tokio::select! race against a cancellation token) kills the child and leaks nothing.

Trait Implementations

impl Clone for LlmEmbedding

fn clone(&self) -> LlmEmbedding

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for LlmEmbedding

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.