LLM bridge: how do I classify a text with a Dense-Armor classifier, and call it from an LLM tool?¶

A Dense-Armor classifier takes a dict of numbers. A text is not a dict of numbers. LLMEmbeddingClassifier closes the gap: you give it a function embed that turns a text into a list of numbers, and it passes those numbers, named e0, e1, and so on, to the classifier it wraps. It also reads and writes JSON strings, which is the form a tool-calling LLM loop uses.

The bridge has no extra package: embed is yours. The examples use the length of the text as the only number.

1. A text in, a label out¶

learn_one(text, label) and predict_one(text) take texts. The four texts below are long or short Italian and English words.

from dense_armor.bridges.llm import LLMEmbeddingClassifier
from dense_armor.utility.online.learn.online_classifiers import OnlineGaussianNB

def embed(t):
    return [len(t)]

clf = LLMEmbeddingClassifier(embed, OnlineGaussianNB())
d = [("buongiorno", "long"), ("buonasera", "long"), ("hello", "short"), ("hi", "short")]
for _ in range(10):
    for t, y in d:
        clf.learn_one(t, y)
print(clf.predict_one("ciao"), clf.predict_one("arrivederci"))

short long

ciao has 4 characters and is closer to the short class; arrivederci has 11 and is closer to the long one.

2. JSON in, JSON out¶

learn_json takes {"text": ..., "label": ...}; predict_json and predict_proba_json take {"text": ...}. Each returns a JSON string, so an agent can call them as tools.

import json
from dense_armor.bridges.llm import LLMEmbeddingClassifier
from dense_armor.utility.online.learn.online_classifiers import OnlineGaussianNB

def embed(t):
    return [len(t)]

c = LLMEmbeddingClassifier(embed, OnlineGaussianNB())
d = [("buongiorno", "long"), ("buonasera", "long"), ("hello", "short"), ("hi", "short")]
for _ in range(10):
    for t, y in d:
        c.learn_json(json.dumps({"text": t, "label": y}))
print(c.predict_json('{"text": "ciao"}'))
print({k: round(v, 3) for k, v in json.loads(c.predict_proba_json('{"text": "ciao"}'))["proba"].items()})

{"prediction": "short"}
{'long': 0.0, 'short': 1.0}

A label that is missing or null raises ValueError, because a label-less sample cannot teach the classifier.

3. A description for the agent¶

describe() returns a JSON-serialisable description of the object, with the input and output schema, for the tool list of an LLM.

from dense_armor.bridges.llm import LLMEmbeddingClassifier
from dense_armor.utility.online.learn.online_classifiers import OnlineGaussianNB

def embed(t):
    return [len(t)]

c = LLMEmbeddingClassifier(embed, OnlineGaussianNB())
print(sorted(c.describe()))

['embedding', 'input_schema', 'methods', 'module', 'name', 'output_schema', 'parameters']

Details¶

References¶

None: the bridge is an adapter, and the three blocks above are the check of its behaviour.

Choices stated explicitly¶

embed is supplied by the caller. The bridge does not call any model, and has no dependency beyond NumPy. The llm extra in pyproject.toml is empty.
Feature names. The vector returned by embed becomes the dict {"e0": ..., "e1": ...}. embed must return a one-dimensional sequence of numbers: a string, None or an array with another shape raises TypeError or ValueError.
Labels in JSON. A string label stays a string; any other label is its JSON text (True becomes "true", 1 becomes "1") in the keys of predict_proba_json.

Measured cost¶

Not measured.