Source Metadata Enrichment Prompt Template

This prompt is designed for a single, high-stakes job: repairing reference lists that are structurally valid but semantically incomplete. The ideal user is a backend engineer, data pipeline operator, or RAG product team member who has a list of citations that passed a basic schema check but is missing required metadata fields—such as DOIs, ISBNs, access dates, or publisher names—that a downstream academic, legal, or compliance system demands. The prompt assumes you already have a list of citation strings and access to a metadata source, such as the Crossref or OpenAlex API, which the model will be instructed to query via a tool. It is not for generating citations from scratch, nor for fixing broken JSON structure; it is for enriching existing, well-formed citations with authoritative external data.

You should use this prompt when your pipeline enforces a strict metadata contract and a citation fails validation because a field like doi or isbn is null. The prompt instructs the model to act as a meticulous metadata librarian: it must query the provided tool for each citation, extract the missing fields, and preserve the original citation text without alteration. It must also produce a structured enrichment report that flags any citation it could not resolve, including the specific fields that remain empty and the reason for the failure. This is a high-precision workflow, not a high-recall one; the prompt is optimized to avoid hallucinating metadata by requiring the model to cite the exact API response that provided each enriched field. If your use case can tolerate best-guess metadata without source grounding, a simpler extraction prompt will suffice.

Do not use this prompt when the input is a block of free text with inline references that haven't been parsed into a structured list. That is an extraction and normalization problem, not an enrichment one. Similarly, avoid this prompt if your metadata source is a static knowledge base rather than a live API; the tool-use instructions assume a real-time query pattern. Finally, this prompt is inappropriate for consumer-grade applications where a missing publication date is an aesthetic issue rather than a compliance blocker. The enrichment report it generates is designed for automated downstream processing and audit logging, not for direct display to end users. Before implementing, ensure your metadata API tool is robust to rate limits and that your evaluation suite includes test cases for ambiguous citations that match multiple records, as this is the most common production failure mode.

This template is designed to be dropped into a post-generation repair loop. It takes a list of citations that have already passed basic format validation but are missing required metadata fields for downstream ingestion systems. The prompt instructs the model to query a provided metadata source, fill gaps where possible, and flag unresolvable enrichment failures rather than hallucinating plausible values. Use this when your reference pipeline requires complete records before citations can be stored, indexed, or displayed to users.

codeCopy
You are a citation metadata enrichment specialist. Your task is to repair incomplete reference records by filling missing metadata fields using the provided source database. You must never invent metadata. If a field cannot be resolved from the source, flag it explicitly.

## INPUT
[INPUT]

## SOURCE METADATA DATABASE
[SOURCE_DATABASE]

## REQUIRED FIELDS
[REQUIRED_FIELDS]

## OUTPUT SCHEMA
[OUTPUT_SCHEMA]

## CONSTRAINTS
[CONSTRAINTS]

## INSTRUCTIONS
1. Parse each citation from [INPUT] and identify which required fields from [REQUIRED_FIELDS] are missing or empty.
2. For each missing field, query [SOURCE_DATABASE] using available identifiers (title, author, year, existing partial identifiers).
3. If a match is found with high confidence, populate the field with the canonical value from the source.
4. If multiple candidates exist, select the best match and note the ambiguity in the `enrichment_notes` field.
5. If no match is found, leave the field as `null` and set the enrichment status to `unresolvable`.
6. For access dates, use the current date in ISO 8601 format unless a specific access date is provided in [INPUT].
7. Normalize all identifiers to their canonical forms (DOI as URL, ISBN-13 format, etc.).
8. Output the complete enriched citation list following [OUTPUT_SCHEMA] exactly.
9. Include a summary section listing all citations that still have unresolvable gaps.

## ENRICHMENT CONFIDENCE LEVELS
- `exact_match`: Identifier or title matched uniquely in source database.
- `high_confidence`: Multiple fields matched, no ambiguity.
- `moderate_confidence`: Partial match, best candidate selected.
- `low_confidence`: Weak match, field populated but should be reviewed.
- `unresolvable`: No match found, field left null.

## EXAMPLE
Input citation: Smith, J. (2020). Machine Learning Applications. Journal of AI Research.
Missing field: DOI
Source query result: Found DOI 10.1234/jair.2020.5678 with exact title match.
Enriched output includes DOI with `exact_match` confidence.

Now process the citations in [INPUT] against [SOURCE_DATABASE].

Adapt this template by replacing the placeholders with your specific inputs. [INPUT] should contain the citation list as structured JSON or a formatted reference block. [SOURCE_DATABASE] should include the metadata records available for lookup—this can be a JSON blob, an API response, or a pre-retrieved set of records from your catalog. [REQUIRED_FIELDS] defines which fields must be present (e.g., ["doi", "isbn", "access_date", "publisher"]). [OUTPUT_SCHEMA] should specify the exact JSON structure you need downstream, including enrichment status and confidence fields. [CONSTRAINTS] can include rules like "never guess a DOI" or "flag missing access dates for human review." After adapting, run this prompt through your eval suite using citations with known missing fields to verify that enrichment is accurate and that unresolvable gaps are flagged rather than fabricated.

This prompt is designed to be called as a post-processing step after a primary generation or extraction pipeline has produced a reference list. The typical integration point is inside a structured output repair loop: the upstream model outputs a list of citations, a schema validator checks for required metadata fields (DOI, ISBN, access date, publisher), and any record missing those fields is routed to this enrichment prompt. The prompt expects a metadata source—either a live API (CrossRef, OpenLibrary, PubMed) or a pre-fetched cache—to be provided in the [METADATA_SOURCE] placeholder. Do not call this prompt without a concrete metadata source; the model cannot invent valid DOIs or ISBNs and will hallucinate if asked to do so.

Wire the prompt into a retry queue with a maximum of two enrichment attempts per record. On the first attempt, pass the full [METADATA_SOURCE] context. If the output still contains unresolvable fields, trigger a second attempt with a narrower [CONSTRAINTS] block that instructs the model to mark those fields as unresolvable rather than guessing. After two failures, escalate the record to a human review queue with the original citation, the attempted enrichments, and the metadata source response logged. Log every enrichment attempt with the input citation, the metadata source queried, the fields enriched, the fields flagged as unresolvable, and the model and prompt version used. This audit trail is essential for compliance-sensitive systems where source provenance must be traceable.

For model choice, prefer a model with strong instruction-following and low hallucination rates on structured extraction tasks—Claude 3.5 Sonnet or GPT-4o are good defaults. Avoid smaller or older models that are more likely to invent plausible-looking DOIs or ISBNs. If using a local or air-gapped deployment, ensure the metadata source is a static database snapshot rather than a live API, and include the snapshot date in the [CONTEXT] so downstream systems can assess freshness. The output schema should be validated programmatically: check that enriched fields match expected regex patterns (DOI: 10.\d{4,}/.+, ISBN: \d{3}-\d-\d{2,6}-\d{2,6}-\d), that no field contains placeholder text like 'N/A' or 'unknown' unless explicitly flagged as unresolvable, and that the enrichment_status field is present and set to one of the allowed enum values (enriched, partial, unresolvable, unchanged).

Before deploying, build a golden test set of 50 citations with known missing metadata and verified correct enrichments. Run the prompt against this set and measure field-level accuracy (did the DOI match?), unresolvable detection rate (did it correctly flag what it couldn't find?), and hallucination rate (did it invent any identifiers not present in the metadata source?). Set a release gate at ≥95% field accuracy and 0% hallucination rate on the golden set. In production, monitor the ratio of enriched to unresolvable outcomes; a sudden spike in unresolvable may indicate a metadata source outage or schema change. Wire these metrics into your existing prompt observability stack alongside token usage and latency.