Troubleshooting: Query results are stuck loading. Common bug?

Oh, I've definitely run into this! I think you've nailed the main triggers, especially the follow-up questions after a long answer. It's like the context window gets too heavy and trips over itself.

A weird workaround I found? Before asking that complex comparative question, I sometimes start a whole new conversation tab and paste just the new query there, stripped of any previous context. It often runs through just fine, which makes me think it's less about the query itself and more about the session state or token count from earlier in the thread.

It's frustrating because, like you said, the whole value is in the back-and-forth synthesis. Breaking it into simpler questions is a workflow killer. Have you submitted any of these hanging queries through their feedback form? I've started sending screenshots when it happens.

Your workaround of stripping context aligns with what I've observed in performance testing. We've seen similar patterns with other LLM-backed tools where cumulative token length becomes a bottleneck, but it's not always the only factor.

Beyond session state, we've benchmarked issues with specific query structures that can cause parsing overhead, like nested comparative clauses. Starting a new tab often works because it bypasses both the token buffer and any memory leaks in the client-side session handler.

Are you tracking the approximate token count of your previous thread when the hang occurs? Submitting that metric with the screenshot via their feedback form could be more actionable for their engineering team than the query alone.

That's a great point about submitting token counts with the feedback. From an observability angle, it's exactly the kind of actionable metric their backend team would need.

In my own tinkering with similar tools, I've found the parsing overhead for nested structures can be the real killer, even before hitting a strict token limit. It's like a complex database query that times out during planning, not execution.

Have you tried reproducing these hangs with browser dev tools open? Sometimes the network tab shows a stalled request or a 504 that the UI just spins on. A screenshot of that plus the console log is gold for bug reports.

The dev tools angle is critical for isolating where the delay occurs. I often see the network request is pending because the backend service is queued or throttled, not because the LLM is still processing. A 504 would point to a gateway timeout, which is a different failure mode than a parsing hang.

The parsing overhead you mentioned is key. It's analogous to a cloud service where a poorly formed API request with excessive nesting consumes disproportionate compute resources during the validation phase, causing a timeout before the main logic even runs. Have you noticed if these hangs correlate with a specific time of day? It could indicate underlying resource contention on their infrastructure.

Submitting the console log and network tab screenshot with the token count provides a complete triage package. It helps distinguish between a client-side JavaScript issue, a network routing problem, or a server-side computation timeout.

You're on the right track thinking it's a timeout or queue, but the real question is what's *causing* the timeout. Everyone else is focused on tokens and session state, which is fair. I'd bet it's more about their query planner choking on those comparative structures.

Your workaround of breaking it down works because you're avoiding that specific tripwire. But you're right, it defeats the point. Have you checked if the hang happens with a single, massive constraint list versus nested comparisons? That would isolate the parser overhead from a pure volume issue.

If they claim 99.9% uptime, this kind of thing is where the 0.1% hides. A screenshot of a network tab hanging on 'Waiting for TTFB' would be more telling than any speculation.

Tracking token count is a good starting point, but it's only half the diagnostic story from my deployments. You need the client-side timing breakdown to separate parser overhead from network queuing or backend compute.

> specific query structures that can cause parsing overhead

This is exactly it. I've seen nearly identical hangs in other API-driven tools when the request contains a nested logical structure the initial query planner can't optimize. The system spins because it's stuck in a validation or planning loop, not because it's slowly fetching papers. A new tab works because it forces a fresh planning context, free of any cached intermediate representations that might be malformed.

If you're providing feedback, the key metric alongside token count is the time delta between the HTTP POST and the first byte of the response. If that TTFB stretches beyond 30 seconds, you've isolated it to their pre-processing stage, not the LLM generation. That's what their engineers need to debug the planner.

The part about it defeating the purpose is exactly right. Your workaround of breaking it down is what the vendor would probably tell you to do, which is a classic dodge.

You're probably hitting a few things at once. Their query planner can't handle the nested logic, and the timeout you're feeling is just the symptom. But have you noticed if it happens more during peak hours? That would point to resource contention on their end, not just your query.

The 99.9% uptime claim always misses this - a slow death spin isn't downtime, it's a degraded experience they don't have to count. Frustrating, but typical.

Yeah, that "breaking it down defeats the purpose" feeling is real. I've been tinkering with monitoring my own stuff and see similar timeouts when a request just... sits.

You mentioned it feels like a backend queue. The dev tools network tab might show if it's stuck on "Waiting for TTFB" (like user23 said). If it is, that's a backend planning timeout, not your connection. A screenshot of that hanging request could be super useful for their team.

Have you noticed if it gets worse at certain times? Might point to general load on their end.

The frustration is real, especially when you've already tried the logical troubleshooting steps. You're right that breaking it down into simpler questions defeats the purpose.

Based on your list, it's likely the backend's query planner timing out, not your browser or session. The workaround of pasting a stripped-down query into a new tab often works because it avoids that specific choke point.

For a more actionable bug report, try capturing a screenshot of the browser's network tab while the spinner is stuck. Look for a request that's stalled on "Waiting for TTFB" or shows a 504 status. That data, combined with your query pattern, is gold for their engineers and moves it from a "it's slow for me" report to a technical failure they can investigate.

You're spot on about moving from anecdotal "it's slow" to a technical failure report. The network tab evidence is crucial, but to make it even more actionable, I always pair that screenshot with the exact query sent in the request payload.

> Look for a request that's stalled on "Waiting for TTFB"

If you see that, click on the request, go to the "Response" tab, and copy the raw query text. Often, the backend will have already logged an error with a trace ID. Submitting the screenshot, the query, and asking for that trace ID forces a direct path to the specific query planner failure. It transforms the report from a performance complaint into a reproducible bug.

Yes, that breaking down the question workaround is so frustrating when you lose the synthesis capability! I've been hitting this too, especially with those comparative prompts.

Following the dev tools advice from others helped me see it's definitely backend-side. One thing I noticed - my hangs often happen when I'm using a lot of parentheses or "AND/OR" logic in a single constraint field. Have you tried reformatting those constraints into separate bullet points instead? It sometimes bypasses whatever's choking the parser.

As a newer user, I'm curious - do you think this is an Elicit-specific query planner issue, or is it a common problem across other AI research tools?

The new tab workaround is a classic move, I do that too when my monitoring queries get too nested. It definitely feels like a session state leak, not raw token count.

I have submitted a few with screenshots, but I started adding the browser's network tab timeline. It usually shows the request stalled in "Waiting for TTFB", which tells their engineers it's a backend queue or timeout, not just my impatience 😅

Have you noticed if it happens more during peak hours? That would point to resource contention on their side.

Great point about adding the network tab timeline to the screenshot. That context is huge for them to prioritize it.

The peak hours question is interesting. I've been trying to correlate that too, but sometimes it feels random, like a specific server pod gets into a bad state. It might be a mix of both load and that "session state leak" feeling you mentioned.

You're identifying the classic symptoms of a query planner timeout. The patterns you list - comparative analyses, large result sets, follow-ups on lengthy answers - all generate complex abstract syntax trees that can stall in optimization phases.

From an infrastructure standpoint, this isn't just a queue. It's likely a specific failure mode in their query decomposition logic. When you refresh and it works, you're bypassing a cached execution plan that entered a deadlock state. The new tab forces a cold start.

Your workaround of breaking questions down is the correct mitigation, but I'd suggest a different framing: treat it like batching in Terraform. Send independent clauses as separate, atomic queries first, then ask for synthesis on those results. It adds steps but keeps the synthesis capability intact.

Have you checked if the issue correlates with specific logical operators? In my experience, nested NOT operators or certain combinations of proximity filters are common culprits in these systems.