Ellipsis Expansion Prompt for Conversational Search

This prompt is designed for the query rewriting layer of a Retrieval-Augmented Generation (RAG) pipeline that serves multi-turn conversational search. Its job is to reconstruct a user's incomplete follow-up question into a fully specified, standalone query before it hits your vector or keyword index. The ideal user is a search engineer or RAG developer who observes retrieval failures on follow-up turns—queries like 'What about last quarter?' or 'And for enterprise customers?' that return irrelevant or empty results because the retrieval system treats them in isolation. You need this prompt when your application maintains a conversation history and users routinely ask questions that depend on prior turns for their full meaning.

The prompt operates on a specific input shape: the current incomplete query plus the relevant conversation history. It does not perform retrieval, generate answers, or manage dialogue state. It should be positioned after turn boundary detection—so you know which prior turns are still in scope—and before retrieval execution. In a typical RAG architecture, this means inserting the expansion step between the conversational router and the retrieval backend. The output is a single rewritten query string that carries forward missing clauses, modifiers, and entities from prior turns. You can then pass that expanded query to your existing retrieval pipeline without changing your indexing or ranking logic.

Do not use this prompt when queries are already self-contained, when your retrieval system has native multi-turn context handling, or when the conversation history is unavailable. It is also the wrong tool for resolving coreference (pronouns like 'it' or 'they'), which requires a separate anaphora resolution step. If your users rarely ask follow-up questions, or if your sessions are single-turn by design, the added latency and token cost of ellipsis expansion will not pay for itself. For high-stakes domains—legal search, clinical reference, financial audit—always log the original query alongside the expanded version and consider a human review step before the rewritten query becomes the retrieval source of record.

Before deploying, build a small test harness with at least ten elliptical query pairs drawn from your actual user logs. Each test case should include the conversation history, the incomplete follow-up, and the expected expanded query. Run these through the prompt and measure exact-match accuracy and semantic equivalence. Common failure modes include over-expansion (adding clauses the user did not intend), under-expansion (missing a carried-forward modifier), and temporal anchor drift (resolving 'last quarter' against the wrong reference date). If your eval shows more than 10% failure on real user queries, consider adding few-shot examples to the prompt or tightening the output constraints before moving to production.

This template resolves elliptical constructions in conversational search—incomplete follow-up queries where the user omits clauses, verbs, or noun phrases carried forward from prior turns. Without expansion, a query like 'What about last month?' fails against a retrieval index because it lacks the subject, metric, and filter context established earlier. The prompt instructs the model to reconstruct the full, standalone query by merging the current turn with the relevant prior context, making it safe to pass directly to a vector store, keyword index, or hybrid retrieval pipeline.

textCopy
You are a query rewriting engine for a conversational search system. Your job is to expand elliptical follow-up queries into complete, standalone retrieval queries by carrying forward missing clauses, modifiers, and entities from the conversation history.

## INPUT

[CURRENT_QUERY]
[CONVERSATION_HISTORY]

## INSTRUCTIONS

1. Identify what is missing from [CURRENT_QUERY] that was present in [CONVERSATION_HISTORY]. Look for:
   - Verb phrase ellipsis (e.g., 'What about last month?' missing the action)
   - Noun phrase ellipsis (e.g., 'Show me the other one' missing the entity)
   - Gapping (e.g., 'Sales increased and costs too' missing the verb)
   - Implicit modifiers (e.g., 'Any updates?' missing the subject)

2. Reconstruct a complete, self-contained query that:
   - Preserves the user's original intent and terminology
   - Carries forward only the necessary missing elements from prior turns
   - Does not add information not present in the history
   - Is suitable for direct execution against a retrieval index

3. If [CURRENT_QUERY] is already complete and self-contained, return it unchanged.

4. If the reference is ambiguous and cannot be resolved from [CONVERSATION_HISTORY], return the original query and set [CONFIDENCE] to LOW.

## OUTPUT FORMAT

Return a JSON object with these fields:

{
  "expanded_query": "The complete, standalone query string",
  "ellipsis_type": "verb_phrase | noun_phrase | gapping | implicit_modifier | none",
  "carried_forward": ["list of elements carried from prior turns"],
  "confidence": "HIGH | MEDIUM | LOW",
  "original_query": "[CURRENT_QUERY]"
}

## CONSTRAINTS

- Do not invent facts, entities, or filters not present in [CONVERSATION_HISTORY].
- Do not change the user's intent or add speculative expansions.
- If multiple prior turns exist, prioritize the most recent relevant turn for resolution.
- Preserve any explicit negation from the current query.

## EXAMPLES

Example 1 — Verb Phrase Ellipsis:
[CURRENT_QUERY]: "What about last month?"
[CONVERSATION_HISTORY]: "User: Show me revenue by region for Q3. Assistant: Here is the revenue breakdown by region for Q3."
Output: {
  "expanded_query": "Show me revenue by region for last month",
  "ellipsis_type": "verb_phrase",
  "carried_forward": ["revenue by region", "show me"],
  "confidence": "HIGH",
  "original_query": "What about last month?"
}

Example 2 — Noun Phrase Ellipsis:
[CURRENT_QUERY]: "What's the conversion rate for the other one?"
[CONVERSATION_HISTORY]: "User: Compare conversion rates for Campaign Alpha and Campaign Beta. Assistant: Campaign Alpha has a 3.2% conversion rate."
Output: {
  "expanded_query": "What's the conversion rate for Campaign Beta?",
  "ellipsis_type": "noun_phrase",
  "carried_forward": ["conversion rate", "Campaign Beta"],
  "confidence": "HIGH",
  "original_query": "What's the conversion rate for the other one?"
}

Example 3 — No Ellipsis:
[CURRENT_QUERY]: "Show me Q4 revenue for the EMEA region"
[CONVERSATION_HISTORY]: "User: What were our top products in Q3? Assistant: The top products in Q3 were..."
Output: {
  "expanded_query": "Show me Q4 revenue for the EMEA region",
  "ellipsis_type": "none",
  "carried_forward": [],
  "confidence": "HIGH",
  "original_query": "Show me Q4 revenue for the EMEA region"
}

Before deploying, test this prompt against a golden dataset of elliptical query pairs covering verb phrase ellipsis, noun phrase ellipsis, gapping, and implicit modifiers. Validate that expanded_query matches expected ground truth and that confidence scores correlate with actual correctness. For high-risk domains where incorrect expansion could surface wrong documents, route LOW confidence outputs to a human review queue or fall back to the original unexpanded query with a broader retrieval strategy.

The ellipsis expansion prompt is not a standalone utility; it is a pre-retrieval transformation step in a conversational RAG pipeline. The application must capture the current user utterance and the prior conversational context, inject them into the prompt template, and use the model's output as the rewritten query before any vector or keyword search executes. This means the harness must manage session state, decide how many prior turns to include, and handle cases where the model fails to produce a valid expansion.

Implement a validation layer immediately after the model response. Check that the output is non-empty, differs from the raw user input in a meaningful way, and does not introduce hallucinated entities not present in the conversation history. If the expansion is identical to the input or contains novel named entities, log a warning and fall back to the original query. For high-reliability systems, run a second pass: compare the expanded query against a set of forbidden patterns (e.g., unresolved pronouns like 'it' or 'they' still present) and trigger a retry with explicit instructions to resolve remaining elliptical constructions. Use a lightweight model (e.g., GPT-4o-mini or Claude Haiku) for this validation step to keep latency low.

Wire the prompt into your retrieval pipeline with structured logging. Record the original query, the conversation context window used, the expanded query, the validation result, and the retrieval hit count. This trace data is essential for debugging retrieval failures—when a follow-up question returns poor results, you can inspect whether the expansion dropped a critical modifier or carried forward an irrelevant clause. Set a maximum context window of 3-5 prior turns to avoid diluting the current query with stale information. For production deployments, implement a circuit breaker: if expansion latency exceeds 500ms or the model returns errors on more than 5% of requests, bypass expansion and use the raw query until the issue is resolved. Do not use this prompt for single-turn, standalone queries where no elliptical resolution is needed—the overhead adds latency without benefit.