k-LLMs adds consensus and field-level likelihoods to OpenAI-compatible extraction.
Run K independent generations on the same prompt, align structured outputs, and reconcile every field into one parsed result with measurable uncertainty.
Built by retab
pip install k-llmsHow it works
Basic usage
k-LLMs keeps the OpenAI call shape and adds consensus with n. In Retab extraction APIs, the equivalent control is n_consensus.
Each candidate is parsed to your schema, then reconciled field-by-field. For arrays of objects, alignment runs first so equivalent items are compared even when order differs across model outputs.
Alignment uses stable-key discovery (or composite keys), fuzzy normalization for near-equivalent keys, multi-pass semantic matching for residuals, and preserves unmatched items as solo rows (zero-loss) before consensus scoring.
from k_llms import KLLMs
from pydantic import BaseModel
client = KLLMs()
class Event(BaseModel):
name: str
date: str
participants: list[str]
response = client.chat.completions.parse(
model="gpt-5.2",
messages=[{"role": "user", "content": "Extract event info: Alice and Bob are going to a science fair on Friday."}],
response_format=Event,
n=4, # Run 4 independent generations for consensus
)
print("Content:", response.choices[0].message.content) # Event dict as a json string
print("Parsed:", response.choices[0].message.parsed) # Event object
print("Likelihoods:", response.likelihoods) # Schema-shaped confidence scores
for i in range(1, len(response.choices)):
print(f"Choice {response.choices[i].index}: {response.choices[i].message.content}")
Why k-llms?
Make non-deterministic extraction operationally reliable
Single-run outputs hide ambiguity. k-LLMs makes disagreement explicit by returning likelihoods that mirror your schema.
You can route high-likelihood fields automatically and send ambiguous fields to validation or human review policies.
Background: this approach follows broader K-LLM patterns explored by teams like Palantir.
Watch: consensus and alignment in practice
A short walkthrough of how parallel candidates become one routed output.