Quick answers to the questions we hear most about StretchGPT, followed by fixes for the most common issues. If you don't find your answer here, the linked deep-dive articles cover each topic in full.
General
What is StretchGPT? StretchSuite's proprietary AI assistant — grounded, multi-turn chat with purpose-built assistant families, plus an operator console for model providers, training, GPU nodes, and safety controls. See Getting Started with StretchGPT.
Is my chat private? Yes. Conversations are isolated to your account and your organization; no one else can see them, and context never crosses conversations.
Which assistant family should I use? Match it to the task — Code for programming, Support for tickets, Research for deep dives, Core for everything else. See Understanding Assistant Families.
Does StretchGPT browse the web? No. It grounds answers only in the StretchSearch index and cites what it finds as [n]. If the index lacks evidence, it answers from general knowledge and says so.
Chat and conversations
Why did my reply come back "blocked"? Your workspace's live runtime isn't activated, or a prerequisite (provider connected, model configured, execution enabled) is missing. The reply lists the exact reasons. See Activating Live Chat.
My message won't send — "too long." Messages are capped at 4,000 characters. Trim or split it.
StretchGPT forgot something from earlier in the thread. It keeps roughly the last 20 completed turns in context. In a long thread, older details age out — restate the key fact or start a focused new conversation.
A conversation vanished from my list. It was likely archived. Archived conversations leave the active list.
How do I rename or delete a conversation? Use the conversation's menu → Rename or Archive/Delete. See Organizing Conversations, Titles, and History.
Grounding and citations
Why are there no citations on my answer? The index had no matching evidence for that question. Ask something more specific, or (operators) check coverage with the retrieval inspector. See Retrieval Context and the Search Bridge.
Can I trust a cited fact? The citation shows the source it came from — open it to verify. Uncited StretchGroup-specific claims deserve a double-check.
What is the "reground" retry? When an answer looks under-supported, StretchGPT can automatically reformulate the question, retrieve better evidence, and re-answer. It only keeps the retry if it's stronger, so it never makes answers worse.
Feedback and training
Does a thumbs-up change the model instantly? No. It creates a review-required training candidate; a human reviews it before anything trains. See Rating Answers and the Feedback Loop.
Why does a thumbs-down ask for a correction? A downvote plus your corrected answer becomes a preference (DPO) example. Without a correction, StretchGPT just notes the rejection — no example is created.
How does data become a model? Approved candidates → exported dataset → training job on a GPU node → trained model → eval → promotion with sign-off. See Reviewing the Training Queue and Managing and Promoting Trained Models.
Administration
Why can't I see a provider's API key? Keys are encrypted at rest and never returned. Lost a key? Rotate it with a fresh one.
My admin action was rejected as unauthorized. Connecting providers, approving training data, and promoting models require step-up authorization with the right scope. Re-authorize and retry.
Promotion refused. Staging and production require a recorded eval run and a sign-off, and you can't skip lifecycle states. Provide what's missing and move through states in order.
A training job never runs. No GPU worker node is polling for that job type, or the training runtime isn't activated. Confirm a node is registered and polling, with a node:poll-scoped credential.
A node's requests are rejected. Likely a stale timestamp (clock drift — the window is 5 minutes) or a reused nonce (each request needs a unique, single-use nonce). See Worker Nodes and the Training Bridge.
Still stuck?
- For activation problems, start with Activating Live Chat and The Safety Envelope and Execution Gate — the blocked reasons in the response are your map.
- For provider issues, see Connecting a Model Provider.
- For training and models, see Reviewing the Training Queue and Managing and Promoting Trained Models.
- If a problem persists after following the relevant guide, contact your StretchSuite administrator with the exact error code or blocked reasons shown — they map directly to the fixes above.
Was this helpful?
Help us improve this article
Use these controls to share whether this answer solved the issue. Feedback helps prioritize updates to StretchSuite Support.