> ## Documentation Index
> Fetch the complete documentation index at: https://docs.pioneer.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Adaptive Inference: automatic continuous retraining

> Pioneer's Adaptive Inference monitors live traffic, collects corrections, retrains a new checkpoint, and promotes it automatically when performance improves.

Most fine-tuned models are static: you train once, deploy, and watch accuracy drift as real-world inputs diverge from your training data. Adaptive Inference breaks that pattern. Pioneer monitors your live inference traffic, identifies high-signal examples, generates training data, fine-tunes a new checkpoint, evaluates it, and helps you promotes it so your model improves in production.

## How it works

Pioneer's Deep Research agent curates a training dataset, a fine-tuning job runs, and the best checkpoint is evaluated before\
anything touches production. You control when a new model version gets promoted. :

<Steps>
  <Step title="You serve inference via Pioneer">
    You call `POST /inference` (or the OpenAI-compatible endpoint) as normal. Inferences are logged automatically and accessible via GET/ inferences.
  </Step>

  <Step title="Pioneer captures high-signal traces">
    As traffic flows through, Pioneer monitors inference results and identifies examples that are ambiguous, low-confidence, or otherwise informative for improving the model. These traces are stored in your inference history and are accessible via `GET /inferences`.
  </Step>

  <Step title="A new checkpoint is trained from your corrections">
    Pioneer uses the high-signal traces — plus any explicit feedback you provide — to generate additional labeled training data. It then fine-tunes a new checkpoint of your model using that data.
  </Step>

  <Step title="Performance is benchmarked before promotion">
    After training completes, Pioneer automatically runs an evaluation against a held-out dataset and reports F1, precision, and recall. Pioneer runs continuous evaluation against the captured traces to measure current model performance. This establishes a baseline before any retraining begins.
  </Step>

  <Step title="The improved checkpoint is promoted">
    The new checkpoint is evaluated against the baseline.Review the evaluation results and deploy the best checkpoint from the deployment page Your `model_id` continues to point to the same endpoint — the underlying model has simply improved. The deployment page highlights the best-performing checkpoint to make this decision easy.
  </Step>
</Steps>

## Submitting feedback

Your explicit corrections are the highest-quality signal for Adaptive Inference. After receiving an inference result, submit feedback using the inference ID:

<CodeGroup>
  ```bash cURL theme={null}
  curl -X POST https://api.pioneer.ai/inferences/YOUR_INFERENCE_ID/feedback \
    -H "X-API-Key: YOUR_API_KEY" \
    -H "Content-Type: application/json" \
    -d '{
      "verdict": "incorrect",  
      "corrected_output": {
        "entities": [
          {"text": "Tim Cook", "label": "person", "start": 10, "end": 18},
          {"text": "Apple", "label": "organization", "start": 0, "end": 5}
        ]
      }
    }'
  ```
</CodeGroup>

Retrieve a list of your past inferences to find IDs for follow-up:

<CodeGroup>
  ```bash cURL theme={null}
  curl "https://api.pioneer.ai/inferences?model_id=YOUR_JOB_ID&limit=50" \
    -H "X-API-Key: YOUR_API_KEY"
  ```
</CodeGroup>

Feedback you submit is incorporated into the next training cycle. The more corrections you provide, the faster the model converges on the behavior you want.

**Enterprise:** For custom retraining schedules, feedback pipelines, or dedicated infrastructure, contact the Pioneer team directly.

<Note>
  Unlimited Adaptive Inference is available on Pro, Research, and Custom (Enterprise) plans. It is not included in the Free plan. Upgrade at [pioneer.ai](https://pioneer.ai) → Settings → Plan, or reach out for enterprise pricing.
</Note>

## Next steps

* [Fine-tune a NER model](/guides/fine-tune-ner) — train your initial model before enabling Adaptive Inference
* [Fine-tune an LLM](/guides/fine-tune-llm) — set up a decoder model for continuous improvement
* [Synthetic Data](/guides/synthetic-data) — generate additional labeled data to supplement production traces