S∀ Snake-All

Contextual ranking engine. Text + context → ranked candidates. Per-factory, on-demand inference.

—
Factories
—
Models on disk
<1ms
Load + Predict
0
In-memory state

Live Factories

Loading...

Endpoints

POST/rank— Rank candidates (text + context)
GET/health— Status
GET/factories— All models
POST/rebuild/{factory}— Retrain (auth)
POST/rebuild— Retrain all (auth)
GET/docs— OpenAPI

POST /rank

{
  "factory_name": "VIP",
  "text": "44.2 clair",
  "item_type": "verre1",
  "client_name": "MARTO ISOLATION",
  "reference": "KOJEMA",
  "line_composition": "442/20/4 G200 FE / WE B / Arg",
  "top_n": 5
}
{
  "factory_name": "VIP",
  "candidates": [
    {"rank": 1, "target": "63442", "score": 0.98, "source": "hot"},
    {"rank": 2, "target": "60442", "score": 0.12, "source": "global"}
  ],
  "hot_available": true,
  "global_available": true,
  "latency_ms": 0.8
}

Architecture

Hot model — trained from verified extractions (recent, contextual, 5 features).
Global model — full article catalog + synonyms (broad coverage, 2 features).
Fuzzy — SequenceMatcher against denomination catalog.
Lookup bypass — numeric text > 500 resolves directly.

Models stored as JSON. Loaded per-request (<0.3ms). No worker memory bloat. S3 archive on retrain.