Concept Expansion Prompt for Vector Retrieval

This prompt is for RAG developers and search engineers who need to improve dense retrieval recall when user queries and target documents use different terminology for the same concept. The job-to-be-done is generating a set of semantically related phrases and abstract concepts from a user query that a vector embedding model can match against document chunks, even when exact keywords fail. The ideal user is someone operating a production retrieval pipeline who has already observed recall gaps caused by vocabulary mismatch—queries that should return relevant documents but don't because the user said 'cost reduction' and the document says 'margin improvement.'

Use this prompt when you have a dense vector index, a known embedding model, and a query that is underspecified or uses language that doesn't match your corpus. The prompt requires a [QUERY] input and optionally accepts [DOMAIN_CONTEXT] to constrain expansion relevance, [CORPUS_SAMPLE] to calibrate against actual document language, and [SIMILARITY_THRESHOLD] to control how far the expansion can drift from the original intent. Do not use this prompt for keyword or sparse retrieval systems—synonym expansion with term weighting is a better fit there. Do not use it when the query is already well-aligned with corpus terminology, as unnecessary expansion introduces noise and reduces precision.

The prompt produces conceptual alternatives, not just synonyms. It generates phrases like 'operational efficiency' from 'cost cutting' or 'customer retention strategy' from 'reduce churn'—transformations that preserve intent while shifting vocabulary. This is distinct from synonym expansion, which stays closer to word-level substitution. The output should include a similarity score per expanded term so you can filter low-confidence expansions before retrieval. In high-stakes domains like healthcare or legal search, always route expanded queries through a human review step or an automated relevance validator before retrieval execution. The next section provides the copy-ready template you can adapt and wire into your retrieval harness.

This template is the core instruction set for a Concept Expansion Prompt. It takes a user query and produces semantically related phrases and abstract concepts that a vector embedding model can match, even when the target documents use different terminology. The prompt is designed to be wrapped in your application's harness, with placeholders for the specific query, domain context, and output constraints you need to enforce.

textCopy
You are a query expansion engine for a dense vector retrieval system. Your job is to take a user's search query and generate a list of conceptual alternatives, related phrases, and abstract ideas that represent the same underlying information need. The goal is to improve recall by finding documents that discuss the same topic using different vocabulary.

## Input Query
[USER_QUERY]

## Domain Context
[DOMAIN_CONTEXT]

## Instructions
1. Analyze the core concept and intent behind the query, not just the surface-level keywords.
2. Generate [NUM_EXPANSIONS] alternative phrases that capture the same concept. These can include:
   - Synonyms for key terms.
   - Broader or narrower conceptual terms.
   - Abstract representations of the idea.
   - Related concepts that a document on this topic would likely discuss.
3. For each alternative, assign a `conceptual_similarity` score from 0.0 to 1.0, where 1.0 is a perfect conceptual match.
4. Do not generate alternatives that change the core intent or introduce contradictory concepts.
5. Do not include the original query in the output list.

## Output Format
Return a valid JSON object with the following structure:
{
  "original_query": "string",
  "expansions": [
    {
      "phrase": "string",
      "conceptual_similarity": float,
      "rationale": "string"
    }
  ]
}

## Constraints
- [CONSTRAINTS]

To adapt this template, replace the square-bracket placeholders with your specific values. [USER_QUERY] is the raw input from the user. [DOMAIN_CONTEXT] should be a brief description of the knowledge base's domain (e.g., 'enterprise cloud infrastructure documentation'). [NUM_EXPANSIONS] controls the breadth of the expansion; start with 5 and tune based on your recall and latency requirements. The [CONSTRAINTS] placeholder is where you add any domain-specific rules, such as 'Do not use competitor product names' or 'Preserve any mentioned error codes exactly as written.' After copying, integrate this template into your application's prompt assembly logic and pair it with the validation harness described in the Implementation Harness section to catch output drift and low-quality expansions before they impact retrieval.

The concept expansion prompt is not a standalone utility; it is a pre-retrieval transformation step that must be integrated into a broader RAG pipeline. The prompt receives a user query and produces a set of semantically related phrases and abstract concepts designed to improve dense retrieval recall. In production, this prompt typically sits between the user input layer and the vector database query step. The harness must handle the prompt's input assembly, output validation, and the decision of whether to use the expanded concepts or fall back to the original query. The primary integration points are: a query pre-processor that formats the prompt with the user's raw query and any available session context, a post-processing validator that checks the output against a defined schema and similarity threshold, and a retrieval dispatcher that merges the original and expanded queries for the vector search call.

Start by defining a strict output contract. The prompt should return a JSON object with an expansions array, where each element contains a phrase string, a similarity_score float between 0 and 1, and a rationale string. The harness must validate this schema before any retrieval occurs. If the model returns malformed JSON, the harness should retry once with a repair prompt that includes the raw output and the expected schema. If validation fails again, log the failure and fall back to the original query. For similarity threshold validation, discard any expansion with a similarity_score below a configurable threshold (start with 0.7 and tune based on your eval results). This prevents low-quality or unrelated concepts from polluting the retrieval step. Log every discarded expansion with the query, the discarded phrase, and the score for offline analysis.

Drift detection is critical for this prompt because concept expansion can easily wander into unrelated semantic territory. Implement a cosine similarity check between the embedding of the original query and the embedding of each expanded phrase. If the similarity drops below a defined drift threshold (e.g., 0.5), flag the expansion for review and exclude it from retrieval. This check catches cases where the model generates concepts that are technically related in some abstract sense but practically useless for your document corpus. For high-stakes domains such as healthcare or legal retrieval, route flagged expansions to a human review queue before they enter the retrieval pipeline. For lower-stakes applications, log the drift events and use them to tune the prompt's [CONSTRAINTS] section or adjust the similarity threshold over time.

Model choice matters here. Concept expansion benefits from models with strong semantic understanding, so prefer models with large context windows and high embedding quality. In latency-sensitive applications, use a smaller, faster model for expansion and accept a higher discard rate. In recall-critical applications, use a more capable model and budget for the extra latency. Always instrument the harness with metrics: expansion latency, validation failure rate, drift detection rate, discard rate, and end-to-end retrieval recall with versus without expansion. These metrics let you measure whether the prompt is actually improving retrieval or just adding cost and latency. If the expansion step consistently fails to improve recall on your eval set, remove it from the pipeline rather than letting it silently degrade performance.

Finally, treat the concept expansion prompt as a versioned artifact. Store the prompt template, the similarity threshold, the drift threshold, and the model identifier together in your prompt registry. When you update the prompt, run your golden query eval set to measure recall changes before deploying. If recall regresses, revert the prompt. The harness should support A/B testing so you can compare retrieval quality with and without expansion in production. Never deploy a concept expansion prompt without a kill switch that falls back to raw query retrieval—this ensures that a prompt failure or model outage does not break the entire retrieval pipeline.