Optimize embedding processor performance - reuse LlamaContext

[malibio]

Overview

The embedding processor currently creates a new LlamaContext for each embedding call, causing significant CPU overhead and underutilizing the GPU. During bulk imports (172 docs), CPU usage spikes to 1200%+ while GPU utilization remains low (~10%).

Problem Statement

Current Behavior:

• Each embedding call creates a new LlamaContext via model.new_context()
• Metal kernel compilation occurs for every context creation
• Context is immediately freed after one embedding
• Results in 1200% CPU during bulk operations with ~10% GPU utilization

Evidence:

• Log shows repeated ggml_metal_init: loaded kernel_* messages (context recreation)
• Log shows ggml_metal_free: deallocating after each embedding
• Activity Monitor shows high CPU, low GPU during embedding processing
• 172 document import took ~54 minutes with heavy CPU load

Code Location: packages/nlp-engine/src/embedding.rs:207-253

// Current: New context created for EACH embedding
fn generate_embedding_llama(&self, text: &str) -> Result<Vec<f32>> {
    // ... 
    let mut ctx = model.new_context(backend, ctx_params)...; // Created every call
    // ... use ctx ...
    // ctx dropped here, Metal resources freed
}

Proposed Solution

Reuse the LlamaContext across multiple embedding calls instead of recreating it each time:

1. Store context as Option<LlamaContext> in EmbeddingService
2. Initialize context lazily on first embedding call
3. Reuse context for subsequent calls
4. Only recreate if context parameters change

Alternative Approach:
Implement batch embedding that processes multiple texts in a single context session.

Acceptance Criteria

• LlamaContext is reused across embedding calls (not recreated each time)
• GPU utilization increases during embedding processing
• CPU usage decreases during bulk embedding operations
• Embedding throughput improves (measure docs/second)
• No regression in embedding quality or correctness
• Tests pass (bun run test:all)
• Code passes bun run quality:fix

Technical Specifications

Reference Files

• embedding.rs: packages/nlp-engine/src/embedding.rs - Main embedding service
• embedding_processor.rs: packages/core/src/services/embedding_processor.rs - Background processor
• embedding_service.rs: packages/core/src/services/embedding_service.rs - Node embedding service

Key Changes

Option A: Persistent Context

pub struct EmbeddingService {
    // ... existing fields ...
    #[cfg(feature = "embedding-service")]
    ctx: Option<LlamaContext<'static>>, // Reusable context
}

Option B: Batch Processing

pub fn generate_batch(&self, texts: Vec<&str>) -> Result<Vec<Vec<f32>>> {
    // Create single context
    let ctx = self.get_or_create_context()?;
    // Process all texts in one session
    texts.iter().map(|t| self.embed_with_context(&ctx, t)).collect()
}

Considerations

• Context lifetime management with llama.cpp backend
• Thread safety if context is shared
• Memory usage of persistent context vs recreation overhead
• Graceful handling if context becomes invalid

Dependencies

• None

Related Issues

• Migrate project documentation to NodeSpace collections #770 (discovered during semantic search benchmark)

Non-Goals

• Changing the embedding model itself
• Modifying the debounce/queue system
• Changing the root-aggregate embedding model