llm-structured-confidence 0.1.0

Extract per-field confidence scores from LLM structured JSON outputs using token-level log-probabilities.

llm-structured-confidence

Extract per-field confidence scores from LLM structured JSON outputs using token-level log-probabilities.

The Problem • Installation • Quick Start • Features • API Reference • Supported Providers

---

Designed for Structured Outputs — available in OpenAI, Gemini, and other providers. Works with any JSON schema, but ideal for ENUM-based classification where the model picks from a fixed set of values.

We recommend litellm as a unified interface for calling any provider with structured output and logprobs.

The Problem

When an LLM returns structured JSON with logprobs, tokens don't align with field values. A token like ":" can merge a colon, a quote, and part of the value — all with one logprob.

Token          Logprob     What it contains
──────────────────────────────────────────────
'{"'           -0.006      { and opening "
'category'      0.000      the key
'":"'          -0.200      closing ", colon, opening "  ← structural, NOT the value
'health'       -0.168      ← actual value content
' and'          0.000      ← actual value content
' wellness'     0.000      ← actual value content
'"}'            0.000      closing " and }

Naively summing all overlapping tokens gives 69% instead of the correct 84.5%.

This library parses the JSON precisely, strips structural tokens, and computes confidence using only the tokens that carry actual value content.

Installation

pip install llm-structured-confidence

Import path:

from llm_structured_confidence import extract_field_logprobs

Or from source:

git clone https://github.com/rodolfonobrega/llm-structured-confidence.git
cd llm-structured-confidence
pip install -e ".[dev]"

Quick Start

import litellm
from llm_structured_confidence import extract_field_logprobs

response = litellm.completion(
    model="gpt-4.1-mini",  # or any provider: "vertex_ai/gemini-2.5-flash", etc.
    messages=[
        {"role": "system", "content": "Classify this text."},
        {"role": "user", "content": "Morning yoga and meditation session"},
    ],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "classification",
            "strict": True,
            "schema": {
                "type": "object",
                "properties": {
                    "category": {
                        "type": "string",
                        "enum": ["sports", "health and wellness", "technology"],
                    }
                },
                "required": ["category"],
                "additionalProperties": False,
            },
        },
    },
    logprobs=True,
    top_logprobs=5,
)

result = extract_field_logprobs(response, field="category")

for value, fl in result.items():
    print(f"{value}: {fl.mean_nonzero_probability:.2%}")  # health and wellness: 84.51%

Features

Three confidence metrics

Metric	Formula	Best for
joint_probability	exp(sum(logprobs))	Strictest — literal sequence probability
mean_probability	exp(mean(logprobs))	General-purpose — fair across token counts
mean_nonzero_probability	exp(mean(logprobs ≠ 0))	ENUM classification — ignores deterministic tokens

[!TIP] With ENUMs, only the first token carries real uncertainty — the rest are forced by the constraint. mean_nonzero_probability filters those out, giving you the model's true confidence regardless of category name length.

Scalar fields

result = extract_field_logprobs(response, field="category")
fl = result["health and wellness"]

Array fields (batch classification)

# {"categories": ["health and wellness", "sports", "technology"]}
result = extract_field_logprobs(response, field="categories")
for value, fl in result.items():
    print(f"{value}: {fl.joint_probability:.2%}")

Pydantic auto-detection

Pass the Pydantic model you used for structured output — the library finds Enum, list[Enum], and Literal fields automatically.

from enum import Enum
from pydantic import BaseModel

class CategoryEnum(str, Enum):
    health_and_wellness = "health and wellness"
    sports = "sports"

class Classification(BaseModel):
    category: CategoryEnum

result = extract_field_logprobs(response, model=Classification)

google-genai native support

Pass a google.genai.GenerateContentResponse directly — converted internally using the same logic as litellm's Vertex AI adapter.

from google import genai
from google.genai import types

client = genai.Client(vertexai=True, project="my-project", location="global")
response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents=[...],
    config=types.GenerateContentConfig(response_logprobs=True, logprobs=5),
)

result = extract_field_logprobs(response, field="category")  # same interface

Token inspection

for token in fl.tokens:
    print(f"  {token.token!r:20s}  logprob={token.logprob:.4f}  prob={token.probability:.2%}")
# 'health'             logprob=-0.1683  prob=84.51%
# ' and'               logprob= 0.0000  prob=100.00%
# ' wellness'          logprob= 0.0000  prob=100.00%

Structural tokens (":", "}", {") are never included.

Top alternatives

for alt in fl.top_logprobs:
    print(f"  {alt.token!r:20s}  prob={alt.probability:.2%}")
# 'health'             prob=84.51%
# 'tech'               prob=15.47%    ← "technology"
# 'sport'              prob=0.01%     ← "sports"

Understanding the Metrics

Why log-probabilities? (numerical stability)

Multiplying many small probabilities causes underflow. Logarithms convert multiplication to addition:

log(A × B) = log(A) + log(B)

So instead of P("health") × P(" and") × P(" wellness") = 0.845, we compute sum(logprobs) = -0.168 and convert back: exp(-0.168) = 0.845.

Why mean_nonzero matters for ENUMs

With the enum ["health and wellness", "sports", "technology"], once the model generates "health", the remaining tokens are forced (logprob = 0). The regular mean gets diluted by those zeros:

"health and wellness" (3 tokens): mean = (-0.168 + 0 + 0) / 3 → 94.6%  ← inflated
"technology" (2 tokens):          mean = (-0.088 + 0) / 2     → 95.7%  ← inflated differently

Longer names get more dilution. mean_nonzero fixes this by averaging only tokens where the model had a choice:

"health and wellness": mean_nonzero = -0.168 / 1 → 84.5%  ← real confidence
"technology":          mean_nonzero = -0.088 / 1 → 91.6%  ← real confidence

API Reference

extract_field_logprobs(response, *, field=None, model=None)

Parameter	Type	Description
response	Any	litellm.ModelResponse, openai.ChatCompletion, or google.genai.GenerateContentResponse with logprobs
field	str | None	JSON field name (e.g. "category"). Takes precedence over model.
model	type | None	Pydantic model — auto-detects Enum/list[Enum]/Literal fields

Returns dict[str, FieldLogprob] — maps each value (as string) to its metrics.

Precedence: field > model > all top-level fields.

FieldLogprob

Attribute	Type	Description
value	Any	The parsed value
tokens	list[TokenInfo]	Tokens included in the calculation
joint_logprob	float	Sum of all token logprobs
joint_probability	float	exp(joint_logprob)
mean_logprob	float	Mean of all token logprobs
mean_probability	float	exp(mean_logprob)
mean_nonzero_logprob	float | None	Mean of logprobs where logprob ≠ 0 (or 0.0 if all zero)
mean_nonzero_probability	float | None	exp(mean_nonzero_logprob) (or 1.0 if all zero)
top_logprobs	list[TopAlternative]	Alternatives from the first uncertain token

TokenInfo

Attribute	Type	Description
token	str	Token text
logprob	float	Log-probability
probability	float	exp(logprob) — property
char_start / char_end	int	Position in the JSON string

TopAlternative

Attribute	Type	Description
token	str	Alternative token text
logprob	float	Its log-probability
probability	float	exp(logprob) — property

How It Works

1. Normalize — detect litellm/OpenAI or google-genai format, convert to common (content, tokens) representation
2. Parse — feed JSON to a Lark LALR parser with position tracking
3. Strip quotes — shrink string ranges by 1 on each side to exclude "
4. Overlap — include only tokens whose character span overlaps the value range
5. Metrics — compute the three logprob metrics from included tokens

JSON:   {"category":"health and wellness"}
                    ^^^^^^^^^^^^^^^^^^^^
                    value range [13, 31) ← quotes stripped

Tokens:  '{"'  'category'  '":"'  'health'  ' and'  ' wellness'  '"}'
          ↑                  ↑                                      ↑
       excluded           excluded                               excluded

Included: 'health' + ' and' + ' wellness'  ✓

Supported Providers

Provider	Response type	Logprobs	Structured output docs
litellm (recommended)	ModelResponse	logprobs=True, top_logprobs=5	JSON mode
OpenAI	ChatCompletion	logprobs=True, top_logprobs=5	Structured Outputs
google-genai	GenerateContentResponse	response_logprobs=True, logprobs=5	Structured output

[!TIP] For classification tasks, consider disabling thinking/reasoning to get cleaner logprobs (no reasoning tokens mixed in). This applies to any model that supports it, but depends on your use case — reasoning may improve accuracy for complex classifications.

• litellm: reasoning_effort="none"
• google-genai: thinking_config=types.ThinkingConfig(thinking_budget=0)

Lower-level API

For custom workflows, internal modules are available:

from llm_structured_confidence._parser import parse_json_spans, build_token_char_ranges, tokens_for_span
from llm_structured_confidence._converter import normalize_response

parsed = parse_json_spans('{"category": "sports", "count": 2}')
# parsed["category"] → _ValueSpan(value="sports", char_start=15, char_end=21)

norm = normalize_response(response)
# norm.content → JSON string, norm.tokens → list of NormalizedToken

[!NOTE] These are underscore-prefixed internal APIs that may change in minor releases. Prefer extract_field_logprobs when possible.

Using with AI Agents

The AGENTS.md file contains a compact API reference designed for LLM-based coding agents (Cursor, Copilot, etc.).

Running Tests

# Unit tests (55 tests, no API calls)
pytest llm_structured_confidence/tests/test_unit.py -v

# E2E tests (6 tests, calls Vertex AI)
pytest llm_structured_confidence/tests/test_e2e.py -v -s

# All tests
pytest -v

Publishing

Release automation for PyPI via GitHub Actions is documented in RELEASING.md.