Synonym Expansion with Source Attribution Prompt

This prompt is for search engineers and RAG developers operating in regulated domains where query transformations must be auditable. The core job-to-be-done is expanding a user query with synonyms and related terms to improve retrieval recall, while producing a verifiable chain of provenance for every added term. The ideal user is someone who already has a controlled vocabulary source—a thesaurus, ontology, or internal knowledge base—and needs to prove to an auditor or downstream reviewer why a specific term was added to a retrieval query. Without this prompt, teams either skip expansion in high-stakes domains (hurting recall) or perform un-attributed expansion that fails compliance review.

You should use this prompt when the cost of an incorrect or unexplained expansion is high. Typical scenarios include legal document retrieval, clinical literature search, financial audit evidence gathering, and regulatory filing review. In these contexts, a reviewer must be able to trace every synonym back to an approved source entry. The prompt requires a [SOURCE_VOCABULARY] input—a structured thesaurus, ontology fragment, or term-to-synonym mapping—and will refuse to generate terms that cannot be grounded in that source. It also emits explicit missing_source alerts when the input query contains concepts absent from the provided vocabulary, preventing silent gaps in coverage.

Do not use this prompt for low-latency, high-throughput search where source attribution overhead is unnecessary. If your retrieval pipeline processes thousands of queries per second and every millisecond counts, the attribution generation and validation steps will add unacceptable latency. Similarly, do not use this prompt when no controlled vocabulary source exists; the prompt is designed to refuse ungrounded expansion, so it will produce empty or near-empty results without a source. For general-purpose synonym expansion without audit requirements, use the standard Synonym Expansion Prompt Template for Keyword Search instead. Before implementing, confirm that you have a machine-readable vocabulary source and that your latency budget can absorb the attribution generation step—typically an additional 200-500ms depending on source size and model choice.

This template is designed for regulated domains where every expanded term must be traceable to an approved source. Replace the square-bracket placeholders with your specific thesaurus, ontology, or knowledge base entries before use. The prompt instructs the model to output only expansions that can be attributed to the provided sources, to flag any terms that cannot be sourced, and to include a completeness check. This is not a general-purpose synonym generator; it is a controlled vocabulary expansion tool for audit-ready retrieval pipelines.

textCopy
You are a controlled vocabulary expansion engine for a regulated domain. Your task is to expand the user's query with synonyms and related terms that improve search recall, using ONLY the provided authoritative sources. You must attribute every expansion to a specific source entry and flag any query terms you cannot source.

## INPUT
Query: [QUERY]

## AUTHORITATIVE SOURCES
[SOURCES]

## CONSTRAINTS
- Output ONLY expansions that can be directly attributed to the provided sources.
- For each expansion, cite the exact source entry (e.g., source name, entry ID, or line number).
- If a query term has no match in the sources, list it under "missing_sources" with a note.
- Do not invent synonyms, use general knowledge, or infer related terms.
- Preserve the original query's logical operators, negation, and entity boundaries.
- If the query contains negated terms (e.g., "without X"), do not expand them.

## OUTPUT SCHEMA
Return a JSON object with the following structure:
{
  "original_query": "string",
  "expansions": [
    {
      "original_term": "string",
      "expanded_terms": ["string"],
      "source_attribution": [
        {
          "source_name": "string",
          "entry_id": "string",
          "relationship": "exact_synonym | broader_term | narrower_term | related_term"
        }
      ]
    }
  ],
  "missing_sources": [
    {
      "term": "string",
      "reason": "not_found_in_sources"
    }
  ],
  "completeness_check": {
    "total_query_terms": "number",
    "terms_with_expansions": "number",
    "terms_without_expansions": "number",
    "all_terms_sourced": "boolean"
  }
}

## EXAMPLES
Query: "cardiac event"
Sources: [{"source": "MedThesaurus v3", "entry": "CARD-001", "term": "cardiac event", "synonyms": ["myocardial infarction", "heart attack", "acute coronary syndrome"]}]
Output: {"original_query": "cardiac event", "expansions": [{"original_term": "cardiac event", "expanded_terms": ["myocardial infarction", "heart attack", "acute coronary syndrome"], "source_attribution": [{"source_name": "MedThesaurus v3", "entry_id": "CARD-001", "relationship": "exact_synonym"}]}], "missing_sources": [], "completeness_check": {"total_query_terms": 1, "terms_with_expansions": 1, "terms_without_expansions": 0, "all_terms_sourced": true}}

## RISK LEVEL
[RISK_LEVEL]

To adapt this template, replace [QUERY] with the user's raw input, [SOURCES] with your structured thesaurus or ontology entries (as a JSON array of objects with source name, entry ID, term, and synonyms), and [RISK_LEVEL] with a note about the domain's tolerance for missing expansions (e.g., "High: missing sources must block retrieval" or "Medium: flag and proceed"). The output schema is strict JSON to enable programmatic validation before the expanded terms are injected into a retrieval pipeline. If your sources are large, consider pre-filtering them to relevant entries before passing them into the prompt to stay within context limits. Always validate the output against the schema before use, and log any missing_sources entries for audit review.

This prompt is designed for regulated domains where every term expansion must be traceable to an authoritative source. The implementation harness must enforce that traceability at runtime, not just hope the model complies. The core integration pattern is a pre-retrieval expansion step: the user's query enters the pipeline, the prompt expands it using the provided thesaurus or ontology, and only terms with valid source attribution are forwarded to the retrieval backend. The harness must validate the output schema before any expanded terms touch the search index.

Start by wrapping the prompt call in a strict validation layer. The expected output is a JSON array of objects, each containing term, source, confidence, and relationship_type fields. Before forwarding terms to retrieval, validate that every object has a non-null source field that matches an entry in your provided knowledge base. Reject or quarantine any term where source is null, "unknown", or missing. Log these rejections as missing_source_alert events for audit review. For high-recall use cases, you may allow terms with confidence above a configurable threshold (e.g., 0.9) even with missing sources, but this must be explicitly flagged and reviewed. Implement a circuit breaker: if more than 20% of terms in a batch fail source validation, halt expansion for that query and fall back to the original user query with a logged warning.

Model choice matters here. Use a model with strong instruction-following and structured output capabilities, such as gpt-4o or claude-3-5-sonnet, and set response_format to json_object with the schema enforced. Temperature should be low (0.0–0.2) to minimize hallucinated sources. Implement retries with exponential backoff for schema validation failures, but cap retries at 3 attempts. On the third failure, log the raw output, alert the on-call channel, and fall back to the original query. For auditability, store every expansion request and response in an append-only log with the query, the provided knowledge base version, the expanded terms, and the validation results. This log becomes your evidence trail for compliance reviews.

Human review should be gated on specific triggers: any query where the missing_source_alert count exceeds zero, any batch where the circuit breaker tripped, or any term flagged with relationship_type equal to "contradiction" or "antonym". Route these to a review queue where a domain expert can approve, correct, or reject the expansions before they enter retrieval. Do not silently drop these cases—silent failures in regulated domains become audit findings. Finally, monitor the ratio of expanded queries to original queries over time. A sudden drop in expansion rate may indicate knowledge base drift or prompt degradation, while a spike in missing-source alerts signals that your thesaurus needs updating.