Grounded Answers and Citations

One of StretchGPT's most important features is grounding: when it can find supporting evidence for your question, it bases its answer on that evidence and cites the sources inline. This article explains how grounding works, how to read citations, and what to do when an answer isn't grounded.

What grounding means

When you send a message, StretchGPT quietly runs your question against the StretchSearch index — the suite's search corpus. If it finds relevant evidence, it composes the answer with that evidence attached and instructs itself to cite the sources it used as [1], [2], and so on. The citation markers appear next to the claims they support, and the matching sources are listed with the answer.

Grounding matters because it turns "trust me" into "here's where this came from." For StretchGroup-specific facts — product details, policies, how a feature works — a grounded answer is one you can verify.

How to read a grounded answer

1. Read the answer top to bottom. StretchGPT leads with the conclusion.
2. Watch for bracketed markers like [2] after a sentence — that sentence is supported by source 2.
3. Scroll to the sources attached to the answer to see the underlying material.
4. If a claim has no citation, treat it as general knowledge rather than something drawn from your indexed content.

When StretchGPT can't ground an answer

Grounding is best-effort. If the StretchSearch index doesn't contain evidence for your question, StretchGPT answers from general knowledge and makes clear that's what it's doing — it will not invent a citation or a source URL. A precise "I can't verify that from your sources" is by design preferred over a confident guess.

The reground retry

StretchGPT includes a cognitive loop that can improve a weak first answer. When enabled, if the assistant's first attempt looks under-supported, StretchGPT reformulates your question, retrieves evidence again with the better query, and re-answers using that fresh evidence. If the retry doesn't produce stronger grounding, StretchGPT keeps the original answer — the feature is fail-safe, so it never makes an answer worse. You don't need to do anything to trigger this; it happens automatically when it can help.

A realistic example

You ask:

"What file storage does StretchDrive use and are uploads encrypted?"

StretchGPT retrieves matching material from the index and replies:

"StretchDrive stores objects in an S3-compatible object store [1], and uploads are handled through the platform's storage service [1]. …"

The [1] markers tell you both claims trace to source 1, which is listed below the answer. If the index had nothing on encryption specifically, StretchGPT would answer the storage part with a citation and note that it can't confirm the encryption detail from your sources.

Tips

• Ask specific questions. "What does StretchDrive use for storage?" grounds better than "tell me about StretchDrive."
• Follow the citations before you act on a fact. Grounding shows its work — use it.
• Absence of a citation is a signal. If a StretchGroup-specific claim isn't cited, double-check it.
• Re-ask if the first answer is thin. Even with automatic regrounding, a sharper question often retrieves better evidence.

Troubleshooting

My answers never show citations. The StretchSearch index may not cover your topic, or grounding evidence wasn't found for your phrasing. Try a more specific, keyword-rich question. If no answer in your workspace is ever grounded, ask your administrator whether the StretchSearch retrieval bridge is configured for StretchGPT.

A citation points to something unexpected. Grounding retrieves the closest matching evidence; if it's off-topic, your question may be ambiguous. Rephrase to disambiguate, and the retrieval will target better.

Can I turn grounding off? Grounding is part of how StretchGPT stays accurate; there's no user-facing switch to disable it. If evidence isn't relevant, StretchGPT simply answers from general knowledge and says so.