Coming Soon to PsyFi Practice: See the Sources Behind Every Grounded Answer

Quick answer: PsyFi Practice grounds clinical answers in a curated reference index — clinical practice guidelines plus a growing set of psychology journal feeds. Today that grounding is invisible in the interface: the model consults the references, but you can't see which ones. An upcoming release fixes that with a collapsed Sources footer under grounded answers. Expand it and you see exactly which reference documents informed the answer — title, section, and a link to the original where one exists. No ship date to announce yet; it's coming in an upcoming release.

Clinicians should never take an ungrounded answer on faith

There's a question every clinician should ask of any AI-generated clinical answer: says who?

Large language models are fluent whether or not they're right. That fluency is precisely the danger in a clinical context — a confident, well-structured paragraph about treatment sequencing reads the same whether it's drawn from a published practice guideline or assembled from the model's general training. In most consumer AI tools, you cannot tell the difference, and the tool doesn't help you try. (It's one of the criteria we'd push you to weigh in any evaluation of AI clinical tools.)

Our position is simple: an AI answer you can't verify is an answer you shouldn't act on. Verifiability isn't a nice-to-have for clinical AI — it's the difference between a reference tool and a rumor.

That's the standard we're building toward, and the next step is making PsyFi Practice's existing grounding visible.

What PsyFi Practice already does under the hood

This part is live today, and worth understanding on its own.

When you ask a clinical question, PsyFi Practice doesn't rely on the model's general training alone. Where reference material can help, it retrieves relevant passages from a curated clinical reference index and puts them in front of the model before it answers. The index has two kinds of content:

• Clinical practice guidelines. Published, authoritative treatment guidance — for example, established practice guidelines for PTSD, major depressive disorder, and insomnia.
• Psychology journal feeds. A growing, curated set of established journals, ingested from the publishers' public feeds. This keeps the index current with recent literature. We ingest abstracts and citation metadata — not scraped full-text articles — and journal-derived material is tracked separately from guideline content.

The retrieved passages arrive with their provenance attached — document title, section, and source link — and the model is instructed to lean on that material when it answers. So the grounding is real, and it's already shaping the answers you get.

Here's the gap: you can't see it. The model may reference its sources in how it reasons, but the interface shows you only the finished answer. The receipts exist; the UI keeps them in the back office.

Coming soon: the Sources footer

In an upcoming release, grounded answers will carry a small, collapsed footer: Sources (3) — the count reflecting how many distinct reference documents informed that answer.

It's collapsed by default because most of the time you don't need it — it stays out of the way of the answer itself. When you do need it, one click expands the list, and each entry shows:

• The document title, so you know at a glance what kind of source you're looking at — a practice guideline, a journal item.
• The section, where the source has one — so you're pointed at the relevant part of a long guideline, not just its cover page.
• A link that opens the source in a new tab, when a public URL exists. Sources without a public link — some journal entries, for instance — appear as labeled text.
• A clear journal label on journal-derived entries, so you can immediately distinguish recent-literature material from established guideline content.

The list is deduplicated: if three retrieved passages came from the same guideline, you see that guideline once, not three times.

And one design decision we want to be explicit about: the footer only appears when references were actually consulted. If an answer didn't draw on the index, there's no footer — we will not decorate ungrounded answers with citation theater. The absence of a Sources footer is itself information.

Why this matters more than it looks

A collapsed footer is a small piece of UI. What it changes is the relationship between you and the answer.

It converts trust into verification. Today, trusting a PsyFi Practice answer means trusting our retrieval pipeline sight unseen. With sources visible, you can open the guideline, read the section, and confirm the answer reflects it — the same habit you'd apply to a colleague's claim of "the guidelines say."

It makes answers citable in your own work. If an answer informed by a practice guideline shapes your thinking on a treatment approach, you'll know which guideline — which means you can cite the guideline, not the AI, in your documentation and case consultation.

It keeps us honest. Visible sources cut both ways. If retrieval ever surfaces a source that doesn't actually support the answer, you'll see that too — and we consider that a feature. A grounding pipeline that can't survive its users reading the sources doesn't deserve their trust.

This is the same philosophy behind how we've written about the tool framework: clinical AI earns trust through transparency about what it actually did, not through polish.

When

Soon — it's coming in an upcoming release, and we're not announcing a date. The feature ships when it has passed end-to-end testing against real clinical questions, not before. When it's live, you won't need to do anything: grounded answers will simply start carrying their sources.

This preview is part of a bigger August update to PsyFi Practice — see the full roundup in What's New in PsyFi Practice: August 2026.

PsyFi Practice is HIPAA-aligned clinical documentation software backed by a Business Associate Agreement. For questions about implementation in your practice, contact our onboarding team at [email protected].