← sources

Can someone explain how JEV is different from a simple embeddings model?

source: reddit (reddit:LocalLLaMA)published 2026-10-04 10:41ingested 2026-10-05 02:41

retained text

How is JEV any different from using an embeddings model? I really will appreciate if someone can explain this to me - because I have yet to see the difference.

I'll even give you my JEV server for free! It uses ollama, you install \`ollama pull nomic-embed-text:latest\`.

    % python3 ./jev_embedding.py "How high is the sky?"
    find_phone: 0.38
    volume: 0.41
    calendar: 0.44
    tell_the_time: 0.49
    weather: 0.53
    
    % python3 ./jev_embedding.py "I had this thing on my anus. The doctor burned it off with a laser."
    weather: 0.35
    tell_the_time: 0.36
    calendar: 0.37
    volume: 0.38
    find_phone: 0.43
    
    % python3 ./jev_embedding.py "can you help me locate my phone."
    volume: 0.38
    weather: 0.40
    calendar: 0.43
    tell_the_time: 0.53
    find_phone: 0.89
    
    % python3 ./jev_embedding.py "Hello Cleveland! I can't HEAR you"
    weather: 0.37
    calendar: 0.40
    tell_the_time: 0.44
    find_phone: 0
    volume: 0.56

    #!/usr/bin/env python3
    """
    jev_embedding.py — minimal showcase of the embedding-based intent router,
    excised from jarvis_workflow.py.
    
    Given a phrase on the command line, it embeds the phrase and every example
    utterance (via the local Ollama embedding model), then prints the cosine
    similarity of the phrase to each intent — the raw routing signal — instead of
    running a handler and speaking an answer.
    
        python3 jev_embedding.py "How high is the sky?"
    """
    
    import sys
    import requests
    
    # --- Config (same endpoint/model as jarvis_workflow.py) ---
    OLLAMA_EMBED_URL   = "http://localhost:11434/api/embeddings"
    INTENT_EMBED_MODEL = "nomic-embed-text"
    
    # --- The five cases to detect ---
    # label -> example utterances, matched by similarity.
    INTENTS = {
        "volume": [
            "turn the volume up",
            "make it quieter",
            "set the volume to seven",
        ],
        "tell_the_time": [
            "what time is it",
            "can you tell me the time",
        ],
        "weather": [
            "how's the weather going to be today",
            "will it rain today",
            "do I need a raincoat",
        ],
        "find_phone": [
            "find my phone",
            "where's my phone",
            "ring my phone",
        ],
        "calendar": [
            "when is my next meeting",
            "what's coming up on the calendar tomorrow",
        ],
    }
    def _embed(text):
        """Return a unit-normalised embedding (list of floats) from the Ollama model."""
        r = requests.post(OLLAMA_EMBED_URL,
                          json={"model": INTENT_EMBED_MODEL, "prompt": text},
                          timeout=10)
        vec = r.json().get("embedding")
        if not vec:
            raise RuntimeError("no embedding returned")
        norm = (sum(x * x for x in vec)) ** 0.5 or 1.0
        return [x / norm for x in vec]
    
    
    def _cosine(a, b):
        """Cosine of two unit vectors is their dot product."""
        return sum(x * y for x, y in zip(a, b))
    
    
    def score_intents(text):
        """Best cosine similarity of `text` to each intent's example utterances."""
        q = _embed(text)
        return {label: max(_cosine(q, _embed(ex)) for ex in examples)
                for label, examples in INTENTS.items()}
    
    
    if __name__ == "__main__":
        if len(sys.argv) < 2:
            print('Usage: python3 jev_embedding.py "your phrase"')
            sys.exit(1)
    
        phrase = " ".join(sys.argv[1:])
        scores = score_intents(phrase)
        for label, score in sorted(scores.items(), key=lambda kv: kv[1]):
            print(f"{label}: {score:.2f}")

document fetched 2026-10-05 00:37 · canonical https://reddit.com/r/LocalLLaMA/comments/1wxcn49/can_someone_explain_how_jev_is_different_from_a