Feature: Use proper embedding model for semantic memory

[Roberdan]

Description

Replace the simple hash-based fake embedding with a proper embedding model for semantic memory and RAG functionality.

Current State

Location: src/neural/claude.c:1207

// TODO: Use a proper embedding model (voyage-ai, openai embeddings, etc.)

Currently uses a hash-based fake embedding that doesn't capture semantic meaning.

Options

Option 1: Local MLX Model (Preferred)

• Use the MLX transformer already in codebase
• Requires loading pre-trained weights (see Feature: Load pre-trained MLX embedding weights #2)
• No API costs, works offline
• Best for privacy

Option 2: Voyage AI Embeddings

• High quality embeddings
• API-based (requires key)
• voyage-2 or voyage-code models

Option 3: OpenAI Embeddings

• text-embedding-3-small or text-embedding-3-large
• API-based (requires key)
• Well documented

Requirements

• Generate meaningful semantic vectors
• Support at least 512 dimensions
• Reasonable performance (<100ms per embedding)
• Graceful fallback if service unavailable

Technical Notes

• Memory table already has embedding BLOB column
• Cosine similarity function exists in mlx_embed.m
• Need to update mlx_embed_text() or add alternative

Labels

enhancement, feature-request, embeddings

[Roberdan]

Status Update (2025-12-14)

Status: Placeholder Implementation

Current state in src/neural/claude.c:1220-1257:

• nous_generate_embedding() uses deterministic hash-based pseudo-embeddings
• Same input → same output (consistent)
• Fast local computation, no API calls

Not Semantic: Hash-based approach doesn't capture meaning, only provides stable outputs for testing.

Available Options (documented in code):

1. Local MLX (preferred) - Requires Issue Feature: Load pre-trained MLX embedding weights #2 resolution
2. Voyage AI - voyage-2 or voyage-code models
3. OpenAI - text-embedding-3-small or text-embedding-3-large

Dependencies:

• Issue Feature: Load pre-trained MLX embedding weights #2 for local MLX embeddings, OR
• External API integration for cloud embeddings

Related Work Completed:

• Semantic memory persistence layer: 18 functions in semantic_persistence.c
• /remember and /search commands functional (with placeholder embeddings)

[Roberdan]

Implemented ✅

Proper embedding model is now integrated using OpenAI text-embedding-3-small:

Changes:

1. openai_embed_text() in src/providers/openai.c - calls OpenAI embeddings API
2. nous_generate_embedding() in src/neural/claude.c - uses OpenAI when key available
3. mlx_embed_text() in src/neural/mlx_embed.m - hybrid: OpenAI first, local fallback
4. persistence_save_memory() generates real embeddings when storing memories
5. persistence_search_memories() uses keyword search (semantic search via memory_search tool)

Configuration:

• Set OPENAI_API_KEY environment variable for semantic embeddings
• Without key: hash-based fallback (consistent but not semantic)

Cost:

• ~$0.00002 per 1K tokens (negligible)
• ~768 dimensions per embedding

Decision documented in updated ADR-004: Hybrid Embedding Strategy

Closes #3