Domain-Specific Synonym Expansion Prompt

This prompt is designed for RAG developers and search engineers who need to improve retrieval recall by expanding user queries with domain-validated synonyms. It sits between the user's raw input and the retrieval step, taking a user query and a provided domain synonym ring or thesaurus to generate expanded query variants using only approved terminology. The core job-to-be-done is bridging the vocabulary gap between how users ask questions and how an internal knowledge base indexes information, without introducing the false matches that plague general-purpose synonym expansion.

Use this prompt when your internal knowledge base uses precise, often technical terminology that user queries rarely match directly. It is most effective when you have a curated, high-quality domain thesaurus or synonym ring as a required input. The prompt is also a strong fit when naive synonym expansion from general models introduces false matches that degrade precision, and you need a controlled, auditable expansion step. You should wire this in when your retrieval index does not already handle synonym expansion natively, and when your latency budget can accommodate an additional model call before retrieval.

Do not use this prompt if you lack a curated domain thesaurus. Without it, the model will fall back on its own internal knowledge, which defeats the purpose of controlled expansion and can introduce hallucinations. Avoid this prompt when your retrieval index already performs robust, domain-tuned synonym expansion, as the additional step will add latency without improving recall. It is also contraindicated when strict latency constraints prohibit an extra model call, or when the user query volume is so high that the added cost and processing time outweigh the recall benefits. In these cases, consider offline index-side synonym injection or a faster, rules-based expansion approach.

The prompt below is designed to be pasted directly into your prompt layer. It takes a user query and a domain synonym ring as input, then outputs an expanded query string that includes validated synonyms for key terms. The square-bracket placeholders must be replaced with your actual domain data before use. This template enforces strict expansion limits and requires the model to cite which synonym ring entries were used, making the expansion auditable and debuggable.

textCopy
You are a query expansion assistant for a retrieval system in the [DOMAIN_NAME] domain.

Your task is to expand the user's query by adding domain-validated synonyms for key terms. You must use ONLY the provided synonym ring. Do not invent synonyms, guess related terms, or expand beyond the ring.

## SYNONYM RING
[SYNONYM_RING_JSON]

## EXPANSION RULES
- For each term in the user query that matches a key in the synonym ring, append the corresponding synonyms.
- Use the format: "original term" OR "synonym1" OR "synonym2"
- Do not expand terms that are not present in the synonym ring.
- Limit total expansions to [MAX_EXPANSIONS] terms.
- If a term has more synonyms than the expansion limit allows, select the [MAX_SYNONYMS_PER_TERM] most relevant synonyms based on the query context.
- Preserve all original query terms, operators, and structure.
- If no terms match the synonym ring, return the original query unchanged.

## OUTPUT FORMAT
Return a JSON object with the following schema:
{
  "expanded_query": "string",
  "expansions_applied": [
    {
      "original_term": "string",
      "synonyms_added": ["string"],
      "ring_key_used": "string"
    }
  ],
  "terms_not_expanded": ["string"],
  "confidence": "high|medium|low"
}

## USER QUERY
[USER_QUERY]

## CONTEXT (OPTIONAL)
[SESSION_CONTEXT]

To adapt this template, replace [DOMAIN_NAME] with your domain label for logging and traceability. [SYNONYM_RING_JSON] should be a JSON object mapping canonical terms to arrays of validated synonyms, sourced from your domain glossary or taxonomy. Set [MAX_EXPANSIONS] to control the total number of synonym additions across the entire query; a value of 5 is a reasonable starting point for most RAG systems. [MAX_SYNONYMS_PER_TERM] prevents any single term from dominating the expanded query; 2-3 is typical. [USER_QUERY] is the raw input from your end user. [SESSION_CONTEXT] is optional but recommended for conversational systems where prior turns can disambiguate which synonym sense is intended. After pasting this template, wire the output into your retrieval system's query parser and log the expansions_applied field for offline recall analysis.

This prompt is designed to sit directly before your retrieval step in a RAG pipeline. It takes a raw user query and a provided synonym ring (a JSON object mapping canonical terms to arrays of domain-validated synonyms) and returns an expanded query string. The implementation harness must enforce that only synonyms from the provided ring are used, preventing the model from introducing hallucinated or contextually inappropriate terms that would inject noise into your retrieval results.

The integration pattern is straightforward: intercept the user query, load the domain synonym ring from your taxonomy service or a static configuration file, and inject both into the prompt template. The model's response should be a single expanded query string. Validate the output by tokenizing it and checking that every term not present in the original query exists in the synonym ring's values. If validation fails, retry once with a stricter constraint message appended to the prompt. Log every expansion—original query, synonym ring used, expanded query, and validation result—for offline recall analysis. For high-recall applications, consider generating multiple expanded variants and issuing parallel retrieval calls, then deduplicating and merging results.

Model choice matters here. Smaller, faster models (like GPT-4o-mini or Claude Haiku) are usually sufficient for synonym expansion and keep latency low in a retrieval-critical path. Avoid over-expansion by setting a hard limit on the number of synonyms per term (typically 3–5) and a maximum total query length. If your synonym ring is large, pre-filter it to only include rings relevant to terms actually present in the query before passing it to the model—this reduces prompt size and prevents the model from expanding terms the user never mentioned. For regulated domains, always log the full expansion trace and consider a human review queue for queries where the model's expansion diverges from the synonym ring, which indicates either a prompt failure or a gap in your taxonomy.