Few-Shot Example Provenance Tag Prompt

This prompt is for prompt engineers and AI operations teams who manage libraries of few-shot examples. When you inject examples into prompts at runtime, you need to know which examples were selected, where they came from, and why they were chosen. Without provenance tags, debugging model behavior becomes guesswork. This prompt takes a batch of few-shot examples and annotates each one with structured metadata: source dataset identifier, selection criteria applied, example version, and a content hash. The tagged output lets you audit exactly which examples influenced a model response, compare example sets across prompt versions, and detect stale or unversioned examples before they cause drift.

Use this prompt when your example library has grown beyond a handful of hand-picked samples and you need production-grade traceability. It is designed for teams that assemble prompts programmatically and need to log which examples were injected into each request. The ideal user is an AI engineer or operator who already has a registry of few-shot examples with known source datasets and version identifiers. You should have the raw examples, their source metadata, and the selection criteria (such as embedding similarity, keyword match, or manual curation rules) available as inputs before running this prompt.

Do not use this prompt for generating new examples or evaluating example quality. It is a tagging and audit tool, not a curation engine. If your examples lack source dataset identifiers or version stamps, you need to backfill that metadata first—this prompt annotates existing provenance data onto examples, it does not invent it. Also avoid using this prompt when you only have one or two examples; the overhead of structured tagging only pays off when you need to trace selection decisions across a larger library. For high-risk domains where example provenance has regulatory implications, always pair this prompt's output with a human review step to verify that the tagged metadata matches your system of record.

This template injects provenance metadata into every few-shot example in your prompt assembly pipeline. Each example receives a structured tag containing its source dataset identifier, the selection criteria used to retrieve it, and a version stamp. This allows operations teams to audit which examples influenced a model response and to trace unexpected behavior back to specific example sources or selection logic changes.

textCopy
You are an AI assistant. The following few-shot examples are provided to guide your response style and behavior.

Each example is wrapped in <example> tags and includes provenance metadata in a <provenance> block. Use the examples to inform your output, but do not repeat the provenance metadata in your response.

--- EXAMPLES ---

<example>
<provenance>
  source_dataset: [SOURCE_DATASET_ID]
  selection_criteria: [SELECTION_CRITERIA]
  example_version: [EXAMPLE_VERSION]
  retrieval_timestamp: [RETRIEVAL_TIMESTAMP]
  retrieval_score: [RETRIEVAL_SCORE]
</provenance>
<input>[EXAMPLE_INPUT]</input>
<ideal_output>[EXAMPLE_IDEAL_OUTPUT]</ideal_output>
</example>

<example>
<provenance>
  source_dataset: [SOURCE_DATASET_ID]
  selection_criteria: [SELECTION_CRITERIA]
  example_version: [EXAMPLE_VERSION]
  retrieval_timestamp: [RETRIEVAL_TIMESTAMP]
  retrieval_score: [RETRIEVAL_SCORE]
</provenance>
<input>[EXAMPLE_INPUT]</input>
<ideal_output>[EXAMPLE_IDEAL_OUTPUT]</ideal_output>
</example>

--- END EXAMPLES ---

Now respond to the following user input, applying the patterns demonstrated in the examples above.

User input: [USER_INPUT]

To adapt this template, replace each square-bracket placeholder with real data from your example library. The [SOURCE_DATASET_ID] should reference a canonical dataset name from your registry. The [SELECTION_CRITERIA] field should capture the retrieval method, similarity threshold, or manual curation rule that selected this example. The [EXAMPLE_VERSION] must increment whenever the example text or metadata changes. For production deployments, validate that every assembled prompt contains provenance tags for all injected examples before sending the request to the model. If any example is missing a source_dataset or example_version field, abort the request and log the assembly failure for investigation.

This prompt is designed to be called programmatically as part of an example ingestion or curation pipeline, not as a one-off chat interaction. The primary integration point is immediately after a new few-shot example is authored or imported from a dataset. The application layer should assemble the prompt by injecting the example content, the source dataset identifier, the selection criteria that caused this example to be chosen, and the current library version. The model call should be treated as a structured extraction task: the output is a set of provenance tags that must be appended to the example's metadata record before the example enters the active library.

The implementation harness must enforce strict validation on the model's output before accepting it. Parse the response against a schema that requires source_dataset, selection_criteria, library_version, and tag_timestamp fields. Reject any response where these fields are missing, null, or fail type checks. Implement a retry loop with a maximum of two retries on validation failure, feeding the validation error message back into the prompt's [CONSTRAINTS] block on the subsequent attempt. If validation fails after retries, route the example to a human review queue with the raw model output and the validation errors attached. Log every attempt—including the assembled prompt, the model response, and the validation result—to your prompt observability system using a trace ID that ties this tagging event to the example's unique identifier.

Model choice matters here: prefer a model with strong structured output support and low latency, as this prompt will be called for every example in the library. Enable structured output mode (JSON mode or function calling with a strict schema) rather than relying on parsing free text. The [OUTPUT_SCHEMA] placeholder should be populated with a JSON Schema definition that includes required fields, enum constraints for selection_criteria categories, and a regex pattern for semantic version strings in library_version. For high-compliance environments, add a post-processing step that checks the source_dataset value against a canonical registry of approved datasets and flags any unrecognized sources. The tagged metadata should be stored alongside the example in your library database, indexed for audit queries, and included in any downstream prompt assembly traces so operators can always answer the question: 'Which examples influenced this output, and where did they come from?'