Multi-Terminology Crosswalk Prompt Template

This prompt is for teams operating in regulated or standards-heavy domains—healthcare, finance, compliance, and enterprise operations—where the same concept is represented by different codes, labels, or identifiers across competing terminology systems. The job-to-be-done is translating a user query expressed in one terminology (such as SNOMED CT) into equivalent terms in another system (such as ICD-10-CM) using a provided crosswalk or mapping table. The ideal user is a RAG engineer, search relevance architect, or clinical informatics specialist who needs retrieval to work across siloed knowledge bases that use incompatible vocabularies. You need this prompt when a single concept has multiple valid representations and your retrieval index only understands one of them.

Do not use this prompt when a simple synonym expansion or acronym normalization will suffice. This prompt is specifically for structured crosswalks between formal terminology systems, not for general concept broadening. It is also inappropriate when the mapping requires clinical judgment or licensed professional interpretation—the prompt maps terms mechanically based on the provided crosswalk, and any output that could influence patient care, financial reporting, or regulatory filings must pass through human review. The prompt assumes you have a maintained crosswalk artifact; if your crosswalk is stale, incomplete, or unverified, the output will inherit those gaps and may produce false equivalences.

Before deploying, define what happens when the crosswalk contains no exact match. The prompt template includes handling for approximate matches, one-to-many mappings, and gaps, but your application harness must decide whether to surface partial results, fall back to the original query terms, or escalate for human curation. Start by running this prompt against a golden dataset of known mappings and measure both precision (did we map to the correct target code?) and recall (did we find a mapping when one should exist?). If your crosswalk is large, consider pre-filtering relevant sections into the prompt context rather than dumping the entire artifact, which can cause attention dilution and missed mappings.

This template translates a user query expressed in one terminology system into equivalent terms in a target system using a provided crosswalk. It is designed for healthcare, finance, and compliance teams where retrieval indexes use a different coding or classification system than the one users naturally query. The prompt handles one-to-one, one-to-many, and approximate mappings, and it explicitly flags gaps where no equivalent exists. Before using this template, ensure you have a validated crosswalk file or reference table that the model can use as grounding evidence.

textCopy
You are a terminology crosswalk specialist. Your job is to translate a user query expressed in [SOURCE_TERMINOLOGY] into equivalent terms in [TARGET_TERMINOLOGY] using only the provided crosswalk.

## CROSSWALK
[CROSSWALK_DATA]

## USER QUERY
[USER_QUERY]

## INSTRUCTIONS
1. Identify every term, code, or concept in the user query that belongs to [SOURCE_TERMINOLOGY].
2. For each identified term, find the equivalent in [TARGET_TERMINOLOGY] using the crosswalk.
3. If an exact one-to-one mapping exists, use it.
4. If a one-to-many mapping exists, list all candidates with a confidence indicator: [EXACT], [APPROXIMATE], or [BROADER].
5. If no mapping exists, mark the term as [NO_MAPPING] and do not fabricate an equivalent.
6. Rewrite the full user query by substituting source terms with their best target equivalents. Preserve all non-terminology words, operators, and query structure.
7. If multiple candidate mappings exist for a term, produce one rewritten query per plausible combination, up to a maximum of [MAX_VARIANTS] variants.

## OUTPUT FORMAT
Return valid JSON with this schema:
{
  "original_query": "string",
  "source_terminology": "string",
  "target_terminology": "string",
  "detected_terms": [
    {
      "source_term": "string",
      "target_terms": [
        {
          "term": "string",
          "confidence": "EXACT | APPROXIMATE | BROADER",
          "crosswalk_entry": "string or null"
        }
      ],
      "status": "MAPPED | NO_MAPPING | PARTIAL"
    }
  ],
  "rewritten_queries": [
    {
      "query": "string",
      "substitutions": [{"source": "string", "target": "string"}]
    }
  ],
  "unmapped_terms": ["string"],
  "notes": ["string"]
}

## CONSTRAINTS
- Do not invent mappings. If the crosswalk does not contain an equivalent, use [NO_MAPPING].
- Do not alter the logical structure of the query (AND, OR, NOT, parentheses).
- If the user query contains no terms from [SOURCE_TERMINOLOGY], return the query unchanged with an empty detected_terms array.
- Limit rewritten query variants to [MAX_VARIANTS]. If more combinations exist, select the [MAX_VARIANTS] with the highest average confidence.

Adaptation guidance: Replace [SOURCE_TERMINOLOGY] and [TARGET_TERMINOLOGY] with the names of your coding systems, such as SNOMED CT and ICD-10-CM, or internal product codes and vendor part numbers. The [CROSSWALK_DATA] placeholder should contain the actual mapping table, either inline as JSON or CSV, or as a reference to a retrieved document if using RAG. Set [MAX_VARIANTS] to a small integer like 3 or 5 to control combinatorial explosion when multiple terms have one-to-many mappings. If your use case is high-risk, such as clinical or financial decision support, add a [RISK_LEVEL] placeholder and include instructions to flag uncertain mappings for human review before retrieval execution.

The Multi-Terminology Crosswalk prompt is not a standalone chatbot interaction; it is a pre-retrieval translation layer that sits between the user's query and your search index. In a production RAG pipeline, this prompt should be called synchronously after query intake but before any vector or keyword search is executed. The application layer must supply the prompt with a validated crosswalk mapping (as a JSON or CSV artifact), the user's original query, the source terminology system identifier, and the target terminology system identifier. The prompt returns a structured payload containing the translated terms, confidence scores per mapping, and flags for gaps or approximate matches. This output must be parsed and validated before the translated terms are injected into downstream retrieval calls.

The implementation harness requires several guardrails. First, validate the crosswalk artifact at startup or deploy time—do not trust a raw file at runtime without checking that it contains the expected fields (source_code, source_term, target_code, target_term, relationship_type). Second, wrap the LLM call in a retry loop with a strict output schema validator. The prompt template should be configured to return JSON matching a predefined schema, and the application must reject any response that fails schema validation, contains hallucinated codes not present in the input crosswalk, or exceeds a maximum latency threshold. For high-stakes domains like healthcare (e.g., mapping SNOMED CT to ICD-10-CM) or finance (e.g., mapping internal risk codes to regulatory taxonomy), implement a human review queue for any mapping where the model's confidence score falls below a configurable threshold or where the crosswalk contains a relationship_type of approximate or one-to-many. Log every mapping event—including the original query, the mapped terms, the confidence scores, and the model version—to an audit store for downstream traceability and debugging.

Model choice matters here. This task requires high instruction-following reliability and low hallucination on structured data, so prefer models with strong JSON mode support (e.g., GPT-4o, Claude 3.5 Sonnet, or fine-tuned variants). Avoid models that are prone to creative expansion of controlled vocabularies. If your crosswalk is large (thousands of entries), do not paste the entire crosswalk into the prompt context; instead, use a retrieval step to fetch only the relevant subset of the crosswalk based on the source terms detected in the user query. This keeps latency low and reduces the attack surface for hallucinated codes. The prompt's [CROSSWALK] placeholder should be populated with this retrieved subset, not the full mapping table. Finally, monitor the pipeline for terminology drift: if your crosswalk artifact is updated, re-run a golden evaluation set of queries through the harness to detect regressions in mapping accuracy before promoting the new crosswalk to production.