# Semantic Search + Preference Learning Investigation

Date: 2026-06-28

## Goal

Add semantic recipe search ("etwas Leichtes für heiße Tage") and preference learning ("wenn warm dann kein Eintopf") to the meal-planning pipeline.

## Embedding Model Requirements

- **Dimensions:** 1024 (user requirement)
- **Language:** Multilingual / German-capable
- **Hosting:** Preferred on norris (same provider as LLM models)

## Norris Embedding Endpoint Investigation

Tested all 20 models listed at `https://ai.noris.de/v1/models` against the `/v1/embeddings` endpoint:

### Models tested via `/v1/embeddings`
| Model | Result |
|-------|--------|
| vllm/bge-reranker-v2-m3 | 404 — provider API error |
| vllm/glm-5-2-nvfp4 | 404 — provider API error |
| vllm/gemma-4-31b-it | 404 — provider API error |
| vllm/qwen3.6-27b-nvfp4 | 404 — provider API error |

### Additional embedding models tested by name
| Model | Result |
|-------|--------|
| vllm/bge-m3 | "no keys found that support model: bge-m3" |
| vllm/multilingual-e5-large | "no keys found that support model" |
| vllm/e5-mistral-7b-instruct | "no keys found that support model" |
| vllm/gte-large | "no keys found that support model" |
| vllm/NV-Embed-v2 | "no keys found that support model" |
| vllm/snowflake-arctic-embed-l-v2.0 | "no keys found that support model" |

### Conclusion
Norris recognizes embedding model names (gateway returns "no keys found" not "model not found"), meaning the Bifrost gateway knows about them but no API key has them activated. The `/v1/embeddings` endpoint returns 404 for all currently-available models (LLM models don't serve embeddings).

**Action needed:** User must request norris to activate an embedding model (bge-m3 recommended — 1024 dims, multilingual, strong German performance).

## Local vLLM Check

Checked for local vLLM instances that might serve embeddings:
- `10.0.30.97:8000` (ComfyUI CT) — No route to host (SSH and HTTP both unreachable)
- `10.0.30.102:8000` (OpenWebUI CT) — No response
- `10.0.30.99:8000` (Docker host) — No response
- `localhost:8000/8001` — No response

No local vLLM embedding server found. ComfyUI CT (10.0.30.97) has GPU but was unreachable during investigation.

## Proposed Architecture (pending embedding model availability)

### Database Schema Additions

```sql
-- Add embedding column to recipes table
ALTER TABLE recipes ADD COLUMN embedding JSON NULL COMMENT '1024-dim vector as JSON array';

-- Decision logging
CREATE TABLE meal_decisions (
    id INT AUTO_INCREMENT PRIMARY KEY,
    decided_at DATETIME DEFAULT CURRENT_TIMESTAMP,
    recipe_id VARCHAR(255) NOT NULL,
    decision ENUM('accepted', 'rejected') NOT NULL,
    context JSON COMMENT 'temperature, season, weekday, recent_meals',
    INDEX idx_recipe (recipe_id),
    INDEX idx_decided (decided_at)
);

-- Learned preference rules
CREATE TABLE preference_rules (
    id INT AUTO_INCREMENT PRIMARY KEY,
    rule_pattern VARCHAR(255) NOT NULL COMMENT 'e.g. temp>20=>avoid:eintopf',
    weight FLOAT DEFAULT 1.0 COMMENT 'confidence, grows with confirmations',
    confirmed_count INT DEFAULT 0,
    rejected_count INT DEFAULT 0,
    created_at DATETIME DEFAULT CURRENT_TIMESTAMP,
    UNIQUE KEY uniq_pattern (rule_pattern)
);
```

### Pipeline Flow

```
Meal Planner selects recipes:
  1. SQL filter (rating > 4.0, not used this week)
  2. Preference rules apply (boost/penalize by context)
  3. Semantic search for "mood" query (light, hearty, asian...)
  4. Top-K presented as proposal

User feedback ("nein, kein Eintopf bei dem Wetter"):
  → INSERT INTO meal_decisions (rejected, temp=28°C, category=eintopf)
  → Rule mining detects pattern → INSERT/UPDATE preference_rules
  → Next run: Eintopf auto-downweighted when warm
```

### Embedding Generation Strategy

- Batch-generate embeddings for existing 28k+ recipes (one-time)
- Incrementally embed new recipes as v7 scraper inserts them
- Store as JSON array in `embedding` column (simple, no pgvector/Qdrant dependency)
- Cosine similarity computed in Python at query time (28k × 1024 floats ≈ 115MB RAM, <1s)
- At 100k recipes: ~400MB RAM for full similarity matrix — still feasible in-process

### Alternative: MariaDB VECTOR Type

MariaDB 11.6+ supports native `VECTOR` type with `VEC_DISTANCE` functions. If the Galera cluster upgrades, switch from JSON-stored arrays to native vectors for indexed similarity search. Current MariaDB version on the cluster should be checked before relying on this.