Use a hypothetical document as a retrieval scratchpad

Read the original mechanism

Akshay’s archived explanation introduces HyDE through a generated hypothetical answer. The original paper describes generating an unreal document, encoding it and retrieving real documents from a corpus. The generated text may contain false details. Its role in retrieval does not make it a source, and the paper’s reported experiments do not guarantee improvement on your collection.

Try a controlled question

Use a public documentation question whose answer you can locate by hand. Save the documents returned by a direct-query baseline. Then generate a clearly labeled hypothetical passage and use it only to form the alternative retrieval input. Compare whether each result set contains the answer-bearing passage. Keep the same collection and result count, and record any extra model call or delay.

Prevent evidence confusion

Give the answering stage only verified passages with document identifiers, or clearly separate the scratchpad from those passages. Require it to identify missing support rather than quote the hypothetical text. The supplied prompt is an original experimental aid, not the paper’s prompt or a proven improvement. No retrieval pipeline was executed here. A useful outcome could be discovering that the simpler baseline is sufficient.

Put it to work.

• Keep the direct-query baseline.
• Label hypothetical text.
• Compare retrieved passages.
• Answer from real documents.

Take the template.

Original CooScope text for your own working notes. Fill in the blanks with evidence from your task.

CooScope — Retrieval scratchpad prompt
Original experimental prompt, informed by the HyDE paper.
Question: [question]
Corpus description: [public collection]
Write a short hypothetical passage using vocabulary that might appear in a relevant document. Label it HYPOTHETICAL SEARCH MATERIAL. Do not invent citations, document IDs or quotations. Mark uncertain concepts as possibilities. Return it separately from any factual answer.
Next: compare retrieval against the original query using [fixed result count].
Limit: generated text is not evidence; this prompt is not a validated HyDE implementation.
Evidence context: https://arxiv.org/html/2212.10496

Go to the source.

Original material, with its own scope and publication date. This note is an explanation, not a CooScope performance test.

1. @akshay_pachaar — archived X post (2024-11-25)
2. Gao et al. — Hypothetical Document Embeddings paperLink / source review: 2026-10-15