k-LLMsBuilt by retab

Consensus and field-level likelihoods for OpenAI-style calls.

Run K independent generations, align structured candidates, and return one parsed object with schema-shaped likelihoods.

pip install k-llms

Read docs GitHub

Consensus flow

How it works

Run independent generations, align the structured candidates, and produce one parsed result with likelihoods.

Sample independent outputs

Set n above 1 to generate candidates from the same prompt.

Align structured candidates

Match equivalent objects and list items before comparison.

Reconcile every field

Return parsed data plus likelihoods shaped like the schema.

Basic usage

Add consensus with one parameter.

k-LLMs keeps the OpenAI parse shape. Set n above 1, read the reconciled result from choices[0], and keep alternative candidates in the remaining choices.

Retab extraction APIs expose the same control as n_consensus.

For arrays and nested objects, alignment happens before scoring so order changes do not erase equivalent items.

parsed: Event(name="Science fair", date="Friday", participants=["Alice", "Bob"])
likelihoods: { name: 0.97, date: 0.88, participants: [0.99, 0.99] }
alternatives: choices[1:] preserves each candidate output for audit and review.

Operational reliability

Route the easy fields. Review the uncertain ones.

Single-run extraction hides ambiguity. k-LLMs makes disagreement explicit by returning likelihoods that mirror your schema.

Use those scores as routing inputs: high-confidence fields can flow through automatically, while uncertain fields go to validation or human review.

participants99% / auto

name97% / auto

date88% / review

Start with one call. Ship with measurable confidence.

Start free Get a demo

k-LLMsBuilt by retab

Consensus and field-level likelihoods for OpenAI-style calls.

Run K independent generations, align structured candidates, and return one parsed object with schema-shaped likelihoods.

pip install k-llms

Read docs GitHub

Consensus flow

How it works

Run independent generations, align the structured candidates, and produce one parsed result with likelihoods.

Sample independent outputs

Set n above 1 to generate candidates from the same prompt.

Align structured candidates

Match equivalent objects and list items before comparison.

Reconcile every field

Return parsed data plus likelihoods shaped like the schema.

Basic usage

Add consensus with one parameter.

k-LLMs keeps the OpenAI parse shape. Set n above 1, read the reconciled result from choices[0], and keep alternative candidates in the remaining choices.

Retab extraction APIs expose the same control as n_consensus.

For arrays and nested objects, alignment happens before scoring so order changes do not erase equivalent items.

parsed: Event(name="Science fair", date="Friday", participants=["Alice", "Bob"])
likelihoods: { name: 0.97, date: 0.88, participants: [0.99, 0.99] }
alternatives: choices[1:] preserves each candidate output for audit and review.

from k_llms import KLLMs
from pydantic import BaseModel


class Event(BaseModel):
    name: str
    date: str
    participants: list[str]


client = KLLMs()

response = client.chat.completions.parse(
    model="gpt-5.5",
    messages=[
        {
            "role": "user",
            "content": "Alice and Bob are going to a science fair on Friday.",
        }
    ],
    response_format=Event,
    n=4,
)

event = response.choices[0].message.parsed
scores = response.likelihoods
alternatives = response.choices[1:]

Operational reliability

Route the easy fields. Review the uncertain ones.

Single-run extraction hides ambiguity. k-LLMs makes disagreement explicit by returning likelihoods that mirror your schema.

Use those scores as routing inputs: high-confidence fields can flow through automatically, while uncertain fields go to validation or human review.

participants99% / auto

name97% / auto

date88% / review

Start with one call. Ship with measurable confidence.

Start free Get a demo