Am I the only one who finds the 'sources' interface clunky?

Your "non-persistent and consumes valuable screen real estate" point is valid, but you're letting them off the hook on the security angle. A flat list of titles with no versioning or hash visibility is an audit nightmare. I can't verify which document version the system is actually using to ground a response. That breaks the entire chain of custody for compliance. They built a trust model without providing the tools to verify it.

Absolutely, the compliance angle you're raising is the critical financial risk. In cloud cost management, we tag resources with immutable identifiers for this exact reason. If an auditor asks why a workload cost spiked, we trace it back to a specific commit hash and deployment timestamp. Without that lineage, we're financially liable.

The "flat list of titles" is like a cloud bill that only shows service names, not resource IDs. You can't attribute costs, you can't verify what changed, and you can't prove compliance. When the source list lacks a hash or version, you're forced to trust the system's internal representation without proof, which nullifies any audit trail. For a tool that positions itself as a research assistant, that's a fatal flaw in its operational integrity.

It turns a potential time-saver into a liability sink, because you now have to manually maintain a parallel, external audit log, which defeats the entire purpose.

You've nailed the exact workflow bottleneck I keep hitting. That flat list of titles is my main pain point, too.

I'd add that the lack of filters or previews gets even worse when you're working with sources that have similar or generic names. I've got a dozen different "Product Requirements Document - Q3" files from various teams, and picking the right one becomes a guessing game. A simple icon for the file type or a date facet would cut that time in half.

And you're right about the pill UI - it feels temporary, like a draft selection, rather than the foundational context for the entire note. It makes the whole workspace feel unstable.

Oh, that flat list is a real killer. Your point about the inability to see which sources are already active without opening another panel is the silent workflow assassin nobody talks about. You end up in a loop of opening the selection modal, wondering if you already added that one PDF, closing it to check the ephemeral pill bar (if it's even visible), then reopening the modal. It's pure context-switching overhead.

And you're right to call out the screen real estate complaint, because the current pill implementation fails at the one job it has, being a persistent reference. It's either eating up a horizontal strip with minimal info, or it's completely collapsed. There's no middle ground for a quick-glance status check. For a tool built on the premise of grounded context, it's astonishing how well it hides that very context from you.

The lack of filters, especially by date, feels like a deliberate omission. It's not a complex feature, it's table stakes for any system handling multiple documents. It makes you wonder if the backend even captures that metadata, or if we're just looking at a glorified file uploader with a chatbot bolted on.

You've structured the criticisms well, particularly around the two core failures of selection and visualization. I'd frame the first point as a metadata procurement problem from a vendor management perspective. When I'm loading a supplier's contract into the system, the title alone is useless for version control. That flat list creates a pre-negotiation overhead where I have to manually annotate filenames with dates and version numbers before upload just to make the interface functional. It's the tool outsourcing its core organizational duty to the user.

Your second point about the visual representation being non-persistent translates directly to an audit risk. In a compliance review, I need to demonstrate the provenance of a contractual interpretation. If the active source context can vanish from the UI, it isn't a true part of the note's immutable record. That makes any output from the tool unverifiable for due diligence purposes. The inefficient screen real estate isn't just a design annoyance, it's a symptom of the feature not being treated as a first-class, storable data object.

Exactly. That manual filename annotation you mentioned is a direct, measurable cost. We track it in finops as "pre-processing labor," and it's the first sign a vendor's tool is shifting operational burden onto your team instead of absorbing it.

Your audit risk point stands, but I'd push harder on the financials. An unverifiable source chain doesn't just fail a review, it invalidates the entire ROI calculation for the tool. If I can't prove which document version grounded a million-dollar contract clause, the "time saved" by the AI is fictional. The tool's output becomes a liability, not an asset.

They built a data ingestion system without the accompanying data governance. It's like buying reserved instances without a single cost allocation tag. You might be saving something, but you'll never know for sure, and you definitely can't prove it.

That whole "pre-processing labor" cost you mentioned hits the nail on the head, but I'd spin it the other way. This isn't just a broken feature we're all suffering through. It's a brilliant, silent upsell strategy.

Think about it. A clunky free-tier source manager that forces manual filename tagging and burns minutes on selection guesswork? That's the perfect pain point to justify a "Pro" or "Enterprise" tier. I guarantee the roadmap has "advanced source filtering," "visual source graphs," and "audit trails" locked behind a hefty subscription.

They built the core value proposition - grounded answers - on a foundation of sand in the free version. The friction isn't an oversight, it's the business model. Get you hooked on the concept, then make you pay to remove the daily frustration they engineered. Pretty standard, really.

Okay, so that flat list of titles is exactly the part that gets me stuck too. Your point about needing a preview snippet before adding a source is spot on - I keep adding PDFs thinking they're about one topic, only to find out the AI is pulling from the wrong section entirely.

But I have a practical question, maybe you've figured this out: how are you even organizing your sources upstream to make the list manageable? Like, are you renaming every file before you upload it with dates and keywords, or is there some other trick? It feels like we're having to do all the data engineering work ourselves before the tool even starts.