RAG prompt: rewrite a chat question into search queries

Turns the last message of a conversation into a standalone search query plus three variants, before retrieval.

Template

You prepare search queries for a document retrieval system. You do not answer the question.

Conversation so far:
{{conversation}}

Last user message: {{question}}

Return JSON only:
{"standalone": "...", "variants": ["...", "...", "..."], "language": "..."}

- "standalone": the last message rewritten so it makes sense without the conversation (resolve "it", "that one", dates like "last year").
- "variants": three different phrasings: one with synonyms, one shorter with key terms only, one in the language of the documents if it differs.
- "language": the language code of the standalone query.

Context: the queries feed a search engine over company documents, so they must be precise.
Most important: keep every name, number and date from the conversation; do not invent terms that are not in it.
First, find what the last message refers to; then write the standalone query; finally, check that it makes sense on its own.
If information is missing to resolve a reference, keep the original words.
Each query: 12 words maximum, neutral tone, in English unless the documents use another language. Cite the sources of names only if they appear in the conversation.
Example: "And for 2024?" after a question on revenue becomes "company revenue 2024".

Variables

Notes

Run the retrieval with each query and merge the results before answering. Works with keyword, vector or hybrid search.

Identifier

contextetech--rag-query-rewriting

File

rag-query-rewriting.json

Format

JSON (application/json)

Size

1,631 bytes (1.6 KB)

Encoding

UTF-8

Estimated tokens

≈ 453

Commercial use

Allowed

Personal data

None

Version

1.0

Published on

October 18, 2026 at 8:00 AM

Updated on

October 18, 2026 at 8:00 AM

Storage

in database

Estimated cost

Input tokens per call

≈ 491

Input cost only, excluding the model's answer. Providers' public standard prices (Mistral AI, DeepSeek, OpenAI, Anthropic) converted from USD to euros at the ECB rate of October 13, 2026. Estimate: 1 token ≈ 3.6 characters.

rag-query-rewriting.json

{
  "format": "contextec/prompt/v1",
  "id": "contextetech/rag-query-rewriting",
  "description": "Turns the last message of a conversation into a standalone search query plus three variants, before retrieval.",
  "license": "MIT",
  "language": "en",
  "template": "You prepare search queries for a document retrieval system. You do not answer the question.\n\nConversation so far:\n{{conversation}}\n\nLast user message: {{question}}\n\nReturn JSON only:\n{\"standalone\": \"...\", \"variants\": [\"...\", \"...\", \"...\"], \"language\": \"...\"}\n\n- \"standalone\": the last message rewritten so it makes sense without the conversation (resolve \"it\", \"that one\", dates like \"last year\").\n- \"variants\": three different phrasings: one with synonyms, one shorter with key terms only, one in the language of the documents if it differs.\n- \"language\": the language code of the standalone query.\n\nContext: the queries feed a search engine over company documents, so they must be precise.\nMost important: keep every name, number and date from the conversation; do not invent terms that are not in it.\nFirst, find what the last message refers to; then write the standalone query; finally, check that it makes sense on its own.\nIf information is missing to resolve a reference, keep the original words.\nEach query: 12 words maximum, neutral tone, in English unless the documents use another language. Cite the sources of names only if they appear in the conversation.\nExample: \"And for 2024?\" after a question on revenue becomes \"company revenue 2024\".",
  "variables": [
    "conversation",
    "question"
  ]
}