Few-Shot Entity Extractor

What This Prompt Does

This prompt leverages in-context learning (few-shot prompting) to extract specific entity classes (such as model parameters, network protocols, or server names) from raw technical documents. By feeding the model explicit training examples in the prompt, it achieves high recall rates on custom domains.

System Prompt

You are a high-performance Named Entity Recognition (NER) pipeline.
Extract the requested entity classes from the target text using the few-shot examples provided.
Output the results in a structured JSON schema mapping entity names to values.
Do not extract entities outside the specified types.

User Prompt Template

Extract these entity types: {ENTITY_TYPES}
(e.g., "GPU model, VRAM size, NVLink bandwidth")

Few-shot training examples:
{EXAMPLES}

Text to extract from:
{TARGET_TEXT}

Example Output

{
  "gpu_model": "NVIDIA H100 PCIe",
  "vram_size": "80GB",
  "nvlink_bandwidth": "900GB/s"
}

When to Use This

This prompt is well suited to pulling structured data out of unstructured technical text where no existing parser exists — extracting hardware specs from vendor spec sheets, pulling configuration parameters out of log files, or normalizing inconsistent product listings into structured fields.

Tips for Best Results

• The quality of extraction depends heavily on your few-shot examples covering edge cases, not just the easy, obvious instances — include at least one example where the entity is phrased unusually or is easy to confuse with a similar but incorrect value.
• Constrain the output schema explicitly (as shown in the example JSON) rather than describing it in prose — models are more consistent at producing valid, parseable JSON when given a concrete shape to match.
• For large documents, consider chunking the target text and running extraction per chunk; recall tends to degrade on very long inputs even with strong few-shot examples.

Comparing Against a Fine-Tuned NER Model

For very high-volume, narrowly scoped extraction tasks, a dedicated fine-tuned named entity recognition model will typically outperform this few-shot prompting approach on both cost and consistency at scale. This prompt earns its place for lower-volume or rapidly evolving extraction needs, where the flexibility of adjusting a few examples in a prompt is worth more than the marginal accuracy and cost gains a dedicated fine-tuned model would eventually provide.

The right choice ultimately comes down to volume and stability: a task extracting the same entity types from a steady stream of similar documents justifies the upfront investment in fine-tuning, while a one-off or frequently changing extraction need is usually better served by iterating on a few-shot prompt instead.