Synonym Expansion with Negation Handling Prompt

This prompt is for search relevance engineers and RAG developers who need to expand user queries with synonyms and related terms without corrupting the user's exclusionary intent. The core job-to-be-done is improving recall in keyword and vector search systems while strictly preserving the scope of negated clauses. A naive synonym expander will often add terms like 'cheap' to a query for 'not expensive,' or 'easy' to 'not difficult,' directly contradicting the user's original filter. This prompt solves that by detecting negation boundaries and generating expansions only for the affirmed portions of the query.

Use this prompt when your retrieval pipeline includes a query rewriting stage and you have observed recall failures caused by terminology mismatches in queries containing negations such as 'without,' 'not,' 'except,' 'excluding,' or 'never.' It is particularly valuable in e-commerce search ('shoes without laces'), legal document retrieval ('contracts not governed by California law'), and technical documentation search ('errors except timeout failures'). Do not use this prompt when the user query contains no negation signals, as a simpler synonym expansion prompt will be faster and cheaper. Also avoid it when your downstream index has no synonym support or when latency requirements preclude an additional LLM call in the retrieval path.

The ideal user is a search infrastructure engineer or RAG pipeline owner who controls the query rewriting layer and can integrate structured expansion outputs into a multi-stage retrieval flow. Required context includes the raw user query and, optionally, a domain glossary or controlled vocabulary to constrain expansions. Before deploying, you must pair this prompt with the Synonym Expansion Eval Check Prompt Template to validate that negation boundaries are correctly detected and that no contradictory synonyms leak into negated clauses. In high-stakes domains like legal or medical search, always route outputs through human review before committing to retrieval execution.

This prompt template generates synonym-expanded query variants from a user's original search query while strictly preserving the scope of any negations, exclusions, or prohibitions present in the input. The core challenge it addresses is semantic inversion: when naive synonym expansion replaces a negated term with a synonym that contradicts the exclusionary intent, causing retrieval systems to return documents the user explicitly wanted to avoid. For example, expanding "laptops without overheating issues" to include "thermal problems" as a synonym for "overheating" would invert the negation and retrieve the wrong documents. This template is designed for search relevance engineers and RAG developers who need expanded queries that improve recall without breaking exclusion logic.

textCopy
You are a search query expansion specialist. Your task is to expand a user's original search query with synonyms and related terms to improve retrieval recall, while strictly preserving the scope and intent of any negations, exclusions, or prohibitions present in the query.

## INPUT
Original Query: [ORIGINAL_QUERY]
Domain Context (optional): [DOMAIN_CONTEXT]
Expansion Budget (max terms): [MAX_TERMS]

## NEGATION HANDLING RULES
1. Identify all negation markers in the query: "not", "without", "excluding", "except", "never", "no", "free of", "lacking", "absent", and similar exclusionary language.
2. Determine the exact scope of each negation: which terms, phrases, or concepts are being excluded.
3. When generating synonyms for non-negated terms, proceed normally.
4. **Do NOT generate synonyms for terms inside negation scope.** Adding synonyms for negated terms creates semantic inversion.
5. **Do NOT generate antonyms of negated terms as positive expansions.** The user wants documents without the negated concept, not documents with its opposite.
6. If a negated term has common synonyms that users might search for positively, flag them as "excluded synonyms" in the output.
7. Preserve the original negation markers and their scope in all expanded query variants.

## EXPANSION RULES
- Generate synonyms and related terms only for non-negated concepts.
- Prioritize domain-appropriate terminology when [DOMAIN_CONTEXT] is provided.
- Include both broader and narrower terms where useful for recall.
- Rank expansions by expected retrieval impact, highest first.
- Respect the [MAX_TERMS] budget; if the budget is insufficient for adequate coverage, note this.
- If the query contains no negations, expand normally with synonyms.

## OUTPUT FORMAT
Return a JSON object with this exact schema:
{
  "original_query": "string",
  "negation_analysis": {
    "has_negation": boolean,
    "negation_markers_found": ["string"],
    "negated_terms": ["string"],
    "negation_scope_notes": "string"
  },
  "expanded_terms": [
    {
      "original_term": "string",
      "synonyms": ["string"],
      "is_in_negation_scope": boolean,
      "confidence": 0.0-1.0
    }
  ],
  "excluded_synonyms": [
    {
      "negated_term": "string",
      "synonym_not_added": "string",
      "reason": "inside_negation_scope"
    }
  ],
  "expanded_query_variants": ["string"],
  "warnings": ["string"]
}

## CONSTRAINTS
- Never add synonyms for terms inside negation scope.
- Never convert a negated concept into a positive search term.
- Flag any ambiguity in negation scope as a warning.
- If the query is entirely negated (e.g., "anything except X"), generate only the positive framing of the exception.

Adaptation guidance: Replace [ORIGINAL_QUERY] with the user's raw search string. [DOMAIN_CONTEXT] should describe the knowledge domain (e.g., "enterprise software documentation", "medical literature") to guide synonym selection toward domain-appropriate terminology. Set [MAX_TERMS] to a concrete integer based on your retrieval system's query length limits. The output schema is designed for machine consumption in a retrieval pipeline: expanded_query_variants provides ready-to-use query strings, while negation_analysis and excluded_synonyms provide auditable evidence that negation boundaries were respected. For high-risk domains where retrieval errors have compliance or safety implications, route the warnings field to a human reviewer before executing the expanded queries.

The Synonym Expansion with Negation Handling Prompt is designed to sit between the user's raw query and your retrieval backend. It should be called as a pre-retrieval transformation step, not as a standalone utility. The prompt expects a user query string and an optional domain context, and it returns a structured output containing expanded terms, negation boundaries, and a contradiction risk assessment. This output must be parsed and validated before any terms are injected into your keyword or vector search queries.

To wire this into an application, wrap the prompt in a function that accepts the raw query and a domain context string. The function should call the model with the prompt template, parse the JSON response, and run a validation pass. The validation must check that the negation_boundaries array is non-empty when the query contains explicit negation terms like 'not', 'without', or 'excluding'. It must also verify that no term in the expanded_synonyms array appears in the excluded_terms list. If a contradiction is detected, the harness should either discard the conflicting synonym or escalate for human review depending on the contradiction_risk score. Log the raw query, the expanded output, and the validation result for every call so you can debug retrieval failures later.

For model choice, use a model with strong instruction-following and JSON output capabilities. The prompt includes a strict output schema, so enable structured output mode if your provider supports it. Set a low temperature (0.0–0.2) to reduce variability in negation boundary detection. Implement a retry loop with up to three attempts if the response fails JSON parsing or schema validation. On the final retry, if validation still fails, fall back to using the original query without expansion and log the failure for investigation. Do not silently drop negation constraints to recover from a bad response.

Before deploying, run the Synonym Expansion Eval Check Prompt against a golden set of queries that include negated terms, compound negations, and edge cases like 'all but' or 'everything except'. Measure both the precision of negation boundary detection and the recall of valid synonyms that do not contradict the exclusionary intent. Set a release gate that blocks deployment if the negation boundary F1 score drops below your threshold. After deployment, monitor the contradiction_risk score distribution and the rate of validation failures. A sudden increase in high-risk scores or parse failures often indicates a model behavior change or a shift in user query patterns that requires prompt adjustment.