colip-embeddings/SKILL.md

metadata

name: colip-embeddings
description: >
  Generates multimodal embeddings across olfaction, vision, and language using
  Scentience COLIP models (OVLM, Embeddings Large/Small, OVL Classifier), then
  computes cross-modal similarity for semantic labeling, retrieval, zero-shot
  classification, or experiment comparison. Use when mapping OPU sensor readings
  to natural language descriptions, matching smell episodes to images,
  classifying detected chemicals, or building olfactory search over a dataset of
  prior experiments.
version: 1.0.0
author: kordelfrance
license: Apache 2.0
compatibility: >
  Python: pip install "scentience[models]" (adds torch>=2.0, transformers>=4.30,
  huggingface-hub>=0.16, torchvision>=0.15, Pillow>=9.0, numpy>=1.24). JS: npm
  install scentience + peer dep @huggingface/inference>=2.0. Rust: cargo add
  scentience --features colip (downloads libtorch) or --features colip-system
  (uses system libtorch). Edge/mobile: use COLIP Embeddings Small (exportable to
  Android, iOS, Rust).
metadata:
  domain: olfaction
  tags:
    - embeddings
    - multimodal
    - colip
    - ovlm
    - retrieval
    - classification
    - semantics
    - zero-shot
  models:
    - id: ovlm
      name: OVLM
      architecture: unified-multimodal
      quantization: int8
      target: edge
      description: World's first unified olfaction-vision-language model at the edge
    - id: colip-embeddings-large
      name: COLIP Embeddings Large
      architecture: graph-attention
      target: cloud
      description: >-
        Highest accuracy; use for offline dataset tasks and online high-stakes
        retrieval
    - id: colip-embeddings-small
      name: COLIP Embeddings Small
      architecture: graph-attention
      target: edge
      description: Low-latency; exportable to Android, iOS, Rust via HuggingFace Hub
    - id: ovl-classifier
      name: OVL Classifier
      architecture: graph-attention
      target: edge+cloud
      variants: 2
      description: Returns class probabilities for chemical-to-visual-object links
  hardware:
    - Reconnaisscent
    - Scentinel
    - Olfactory Development Board
  depends_on:
    - ble-device
  docs: https://scentience.github.io/docs-api

Goal

Produce L2-normalized embedding vectors from olfaction sensor readings, images, or text using a COLIP model, then rank candidates by cosine similarity for retrieval, labeling, or classification — while preserving model-version awareness and clearly distinguishing semantic similarity from analytical chemistry.

Instructions

1. Choose a Model Variant

Model	Best for	Deployment	Notes
ovlm	Unified O+V+L tasks, zero-shot labeling	Edge/Mobile	Int8 quantized; on Apple App Store via Sigma
colip-embeddings-large	High-accuracy retrieval, dataset analysis	Cloud	Highest-dimensional embeddings; slowest
colip-embeddings-small	On-device, latency-sensitive pipelines	Edge	Export to Android/iOS/Rust
ovl-classifier	Binary/multi-class compound classification	Edge + Cloud	Returns class probabilities, not vectors

Default recommendation: Start with colip-embeddings-small for robotics and real-time tasks. Switch to colip-embeddings-large for offline dataset work or when accuracy is more important than latency. Use ovl-classifier when you need a hard classification output rather than a similarity ranking.

2. Instantiate the Embedder

# Python — requires pip install "scentience[models]"
from scentience import ScentienceEmbedder

embedder = ScentienceEmbedder(
    model="colip-embeddings-small",   # or "ovlm", "colip-embeddings-large", "ovl-classifier"
    api_key="SCN_..."
)

// JavaScript — requires @huggingface/inference peer dep
import { ScentienceEmbedder } from 'scentience';

const embedder = new ScentienceEmbedder({
  model: "colip-embeddings-small",
  apiKey: "SCN_..."
});

// Rust — Cargo.toml: scentience = { features = ["colip"] }
use scentience::ScentienceEmbedder;
let embedder = ScentienceEmbedder::new("colip-embeddings-small", "SCN_...");

3. Generate Embeddings

All embed() calls return L2-normalized vectors. Pass the correct modality flag.

From an OPU reading (olfaction):

# reading is a dict from the ble-device skill
reading = {"VOC": 0.45, "CO2": 412.0, "NH3": 0.08, "H2S": 0.002, ...}
olf_vec = embedder.embed(modality="olfaction", data=reading)
# Returns: np.ndarray, shape (D,), L2-normalized

From text:

txt_vec = embedder.embed(modality="text", data="freshly cut grass after rain")

From an image:

from PIL import Image
img_vec = embedder.embed(modality="vision", data=Image.open("scene.jpg"))

JavaScript (any modality):

const olfVec = await embedder.embed({ modality: "olfaction", data: reading });
const txtVec = await embedder.embed({ modality: "text", data: "ammonia leak" });

4. Rank Candidates by Cosine Similarity

import numpy as np

def cosine_sim(a: np.ndarray, b: np.ndarray) -> float:
    return float(np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b) + 1e-8))

candidates = [
    {"id": "label_0", "modality": "text", "label": "petroleum / gasoline"},
    {"id": "label_1", "modality": "text", "label": "ammonia / fertilizer"},
    {"id": "label_2", "modality": "text", "label": "fresh cut vegetation"},
]

results = sorted(
    [
        {
            "id": c["id"],
            "label": c["label"],
            "similarity": cosine_sim(
                olf_vec,
                embedder.embed(modality=c["modality"], data=c["label"])
            )
        }
        for c in candidates
    ],
    key=lambda x: x["similarity"],
    reverse=True
)

5. OVL Classifier (Hard Classification)

When the task is binary or multi-class classification rather than retrieval:

from scentience import ScentienceEmbedder

classifier = ScentienceEmbedder(model="ovl-classifier", api_key="SCN_...")
probs = classifier.classify(modality="olfaction", data=reading)
# Returns: dict of {class_label: probability}
# e.g., {"ammonia": 0.83, "methane": 0.09, "voc_mix": 0.05, ...}

Use this when you need a direct answer ("is this ammonia?") rather than a ranked list of semantic matches.

6. Format Output

{
  "model": "colip-embeddings-small",
  "model_version": "0.2.0",
  "query_modality": "olfaction",
  "top_matches": [
    {"id": "label_1", "label": "ammonia / fertilizer", "similarity": 0.84},
    {"id": "label_2", "label": "fresh cut vegetation",  "similarity": 0.61},
    {"id": "label_0", "label": "petroleum / gasoline",  "similarity": 0.33}
  ],
  "semantic_summary": "Query embedding strongly aligned with ammonia/nitrogen compounds. Consistent with NH3=0.08 ppm and elevated VOC. Top match similarity 0.84 with a 0.23 gap to second — treat as confident hypothesis.",
  "confidence": 0.79,
  "note": "Semantic similarity score — not a certified chemical assay. Treat as hypothesis pending corroborating sensor evidence or manual review."
}

Always include model and model_version. Embedding spaces change between model versions; cross-version comparisons are invalid.

Examples

Label a robot smell episode in a warehouse

Scenario: Warehouse inspection robot. Buffer of 10 readings.
          NH3=0.22 ppm (elevated), VOC=0.31 ppm, CO2=420 ppm.

Candidate labels: ["cleaning products", "refrigerant leak",
                   "ammonia coolant system", "diesel exhaust"]

Top match:       "ammonia coolant system"  (similarity 0.88)
Second match:    "cleaning products"        (similarity 0.54)
Gap:             0.34 — confident result

Semantic summary: Elevated NH3 with low VOC/CO2 ratio aligns with ammonia-based
                  refrigerant systems. Recommend inspection of cooling units in sector 3.
Confidence: 0.82

Cross-modal: image query against olfaction database

Input: Image of a fertilizer storage area (vision embedding).
Database: 50 prior olfaction experiment embeddings from ScentNet field trials.

Top match:  Experiment #12 — outdoor soil nitrogen test (similarity 0.76)
Semantic summary: Visual cues of bagged nitrogen fertilizer align with prior
                  high-NH3 olfaction episodes. Cross-modal alignment strong.

Zero-shot classification with OVL Classifier

Input: OPU reading with H2S=0.45 ppm, SO2=0.12 ppm, NH3=0.03 ppm.

OVL Classifier output:
  {"hydrogen_sulfide_source": 0.79, "industrial_exhaust": 0.13, "other": 0.08}

Interpretation: High-confidence H2S-linked source. Confidence 0.79 > 0.65 threshold.

Constraints

• Similarity ≠ chemical identity — COLIP similarity reflects statistical associations from training data, not analytical GC-MS results. Treat high similarity as a strong hypothesis, not a certified identification.
• Always record model name and version — Embedding spaces shift between versions; comparisons across versions are invalid and should be blocked programmatically.
• Normalize before external comparisons — All vectors from ScentienceEmbedder.embed() are L2-normalized. If you supply embeddings from an external system, normalize first.
• Ambiguity threshold — If top-match similarity < 0.65, or if the gap between first and second match is < 0.10, treat the result as ambiguous. Request additional sensor readings or flag for manual review.
• Cross-modal alignment quality varies — Olfaction-to-text alignment is strongest for compound classes present in the ScentNet training corpus. Novel or rare compounds degrade similarity quality.
• OVL Classifier vs. Embeddings — Use ovl-classifier for hard classification (class probabilities). Use colip-embeddings-large/small for nearest-neighbor retrieval and semantic ranking tasks.
• Edge model accuracy tradeoff — COLIP Embeddings Small sacrifices accuracy for size. For safety-critical detection tasks, validate against Large before deploying Small on-device.
• Training data bias — All COLIP models inherit biases from the ScentNet dataset. Performance on out-of-distribution chemical environments (novel compounds, unusual mixtures) is less reliable and requires human validation.