| <!DOCTYPE html> |
| <html lang="en"> |
| <head> |
| <meta charset="UTF-8" /> |
| <meta name="viewport" content="width=device-width, initial-scale=1.0" /> |
| <title>help · rag-psych</title> |
| <link rel="stylesheet" href="/static/styles.css" /> |
| <script src="https://cdn.tailwindcss.com"></script> |
| <script src="https://cdnjs.cloudflare.com/ajax/libs/gsap/3.12.5/gsap.min.js"></script> |
| <script type="importmap"> |
| { "imports": { "three": "https://unpkg.com/three@0.169.0/build/three.module.js" } } |
| </script> |
| </head> |
| <body class="bg-slate-950 text-slate-100 min-h-screen overflow-x-hidden"> |
|
|
| <canvas id="neural-bg" class="fixed inset-0 w-full h-full -z-10 opacity-40"></canvas> |
|
|
| <main class="relative max-w-4xl mx-auto px-6 py-10"> |
|
|
| <nav class="mb-8 flex items-center justify-between text-sm"> |
| <a href="/ui" class="text-slate-400 hover:text-cyan-300 transition-colors"> |
| ← back to search |
| </a> |
| <span class="text-slate-600 uppercase tracking-widest text-xs">help</span> |
| </nav> |
|
|
| <header class="mb-12" id="help-hero"> |
| <h1 class="text-4xl font-light tracking-tight mb-3"> |
| <span class="text-cyan-300">what</span> |
| <span class="text-slate-500">it</span> |
| <span class="text-fuchsia-300">does</span> |
| </h1> |
| <p class="text-slate-400 text-[15px] leading-relaxed"> |
| <strong class="text-slate-200">rag-psych</strong> is a retrieval-augmented |
| question-answering system over a local corpus of psychiatry / |
| mental-health reference material. You type a clinical question; |
| the system finds the most relevant passages in the corpus, has |
| an LLM compose a grounded answer with citations back to those |
| passages, and shows you the supporting passages alongside the |
| answer so you can verify every claim. |
| </p> |
| </header> |
|
|
| |
| <section class="mb-12" id="offers"> |
| <h2 class="text-xs uppercase tracking-widest text-cyan-300 mb-4">what it offers</h2> |
| <div class="grid md:grid-cols-2 gap-3"> |
| <div class="help-card"> |
| <h3 class="text-slate-100 font-medium mb-1">Grounded answers</h3> |
| <p class="text-slate-400 text-sm leading-relaxed"> |
| Every factual claim in the response is followed by a |
| <code class="citation-sample">[chunk_id]</code> citation |
| linking to the exact passage it came from. Click a citation |
| to scroll to and highlight its chunk. |
| </p> |
| </div> |
| <div class="help-card"> |
| <h3 class="text-slate-100 font-medium mb-1">Source transparency</h3> |
| <p class="text-slate-400 text-sm leading-relaxed"> |
| Retrieved passages are shown on the right with their source |
| (clinical notes, research abstracts, or diagnostic references) |
| colour-coded and labelled. No hidden reasoning. |
| </p> |
| </div> |
| <div class="help-card"> |
| <h3 class="text-slate-100 font-medium mb-1">Hallucination detection</h3> |
| <p class="text-slate-400 text-sm leading-relaxed"> |
| Cited IDs that do not appear in the retrieved set are flagged |
| in the answer and in a warning banner. The model does not get |
| to quote things that weren't retrieved. |
| </p> |
| </div> |
| <div class="help-card"> |
| <h3 class="text-slate-100 font-medium mb-1">Insufficient-evidence refusal</h3> |
| <p class="text-slate-400 text-sm leading-relaxed"> |
| When the corpus doesn't contain an answer, the system returns |
| a canonical refusal string rather than inventing one. Off-topic |
| queries trigger this at the retrieval layer with no LLM call. |
| </p> |
| </div> |
| <div class="help-card"> |
| <h3 class="text-slate-100 font-medium mb-1">Negation-aware retrieval</h3> |
| <p class="text-slate-400 text-sm leading-relaxed"> |
| Passages that <em>deny</em> the queried concept ("patient |
| denies suicidal ideation") are filtered out before reaching |
| the answer step, so they're never cited as positive evidence. |
| </p> |
| </div> |
| <div class="help-card"> |
| <h3 class="text-slate-100 font-medium mb-1">Hybrid retrieval</h3> |
| <p class="text-slate-400 text-sm leading-relaxed"> |
| Three retrievers run in parallel — dense semantic search, |
| BM25-style keyword, and literal rare-token matching — then |
| Reciprocal Rank Fusion and a cross-encoder re-score the |
| combined candidate pool. |
| </p> |
| </div> |
| </div> |
| </section> |
|
|
| |
| <section class="mb-12" id="examples"> |
| <h2 class="text-xs uppercase tracking-widest text-cyan-300 mb-4">what to ask</h2> |
| <div class="space-y-3 text-sm"> |
| <div class="help-example"> |
| <span class="help-example-tag">diagnostic</span> |
| <p>criteria for generalized anxiety disorder</p> |
| <p>essential features of post-traumatic stress disorder</p> |
| <p>diagnostic criteria for obsessive compulsive disorder</p> |
| </div> |
| <div class="help-example"> |
| <span class="help-example-tag">clinical scenarios</span> |
| <p>45-year-old female presenting with depressive symptoms and suicidal ideation</p> |
| <p>patient medication list including SSRI for depression</p> |
| </div> |
| <div class="help-example"> |
| <span class="help-example-tag">research</span> |
| <p>cognitive behavioral therapy outcomes for anxiety disorders in adolescents</p> |
| <p>psychosocial interventions for bipolar disorder</p> |
| </div> |
| <div class="help-example"> |
| <span class="help-example-tag">cross-source</span> |
| <p>what does the literature say about the diagnostic criteria for depression</p> |
| <p>how is suicidal ideation assessed clinically and what is its prevalence</p> |
| </div> |
| </div> |
| </section> |
|
|
| |
| <section class="mb-12" id="limits"> |
| <h2 class="text-xs uppercase tracking-widest text-amber-400 mb-4">what it can't do</h2> |
| <ul class="space-y-2 text-sm text-slate-300 leading-relaxed list-disc list-inside"> |
| <li> |
| <strong class="text-slate-100">Medical advice.</strong> Answers |
| are grounded in reference material, not a clinician's judgement. |
| Never use this to make a real diagnostic or treatment decision. |
| </li> |
| <li> |
| <strong class="text-slate-100">Real-time information.</strong> |
| The corpus is a snapshot. It doesn't know about new papers, |
| guidelines, or drug approvals published after ingest time. |
| </li> |
| <li> |
| <strong class="text-slate-100">PHI-sensitive work.</strong> |
| All source data is public or de-identified. Do not paste |
| identifiable patient information into queries. |
| </li> |
| <li> |
| <strong class="text-slate-100">Multi-hop reasoning.</strong> |
| Each query is answered in a single pass. Questions that need |
| separate lookups and then a comparison |
| ("<em>is the criterion for X different in version Y vs Z?</em>") |
| are handled more loosely than direct lookups. |
| </li> |
| <li> |
| <strong class="text-slate-100">Exact-string drug dosing.</strong> |
| If the corpus doesn't contain a literal "<em>drug name + dose</em>" |
| mention, the system may return same-class alternatives rather |
| than the precise dose you asked for. |
| </li> |
| </ul> |
| </section> |
|
|
| |
| <section class="mb-12" id="pipeline"> |
| <h2 class="text-xs uppercase tracking-widest text-cyan-300 mb-4">how a query flows</h2> |
| <ol class="space-y-3 text-sm text-slate-300 leading-relaxed list-decimal list-inside"> |
| <li>Your query is embedded with a clinical-domain sentence |
| encoder, and in parallel tokenised for keyword and rare-token |
| lookups.</li> |
| <li>Three retrievers run against the local vector database and |
| return their top candidates independently.</li> |
| <li>The candidate lists are fused by Reciprocal Rank Fusion, |
| deduplicated, and re-scored by a cross-encoder reranker.</li> |
| <li>A rule-based negation filter drops any surviving passage |
| where the queried concept is denied / ruled out / negative for.</li> |
| <li>If the best remaining passage clears a confidence threshold, |
| the top-k are sent to a language model with a strict system |
| prompt: answer only from these, cite every claim, refuse cleanly |
| if the passages don't support an answer.</li> |
| <li>The response is parsed for citation integrity — any cited |
| ID not in the retrieved set is flagged before the answer is |
| rendered.</li> |
| </ol> |
| </section> |
|
|
| |
| <section class="mb-12" id="roadmap"> |
| <h2 class="text-xs uppercase tracking-widest text-cyan-300 mb-4">what it could offer next</h2> |
| <ul class="space-y-2 text-sm text-slate-300 leading-relaxed list-disc list-inside"> |
| <li><strong class="text-slate-100">Per-source retrieval balance.</strong> |
| Currently one source can crowd out others when a query spans |
| topics. Fetching top-K from each source independently before |
| fusion would keep all three voices in the final answer.</li> |
| <li><strong class="text-slate-100">Stronger reranker.</strong> |
| The current cross-encoder is a general-purpose MS-MARCO model. |
| Swapping to a clinical-tuned reranker would reduce the |
| "case-study" bias observed in the eval set.</li> |
| <li><strong class="text-slate-100">Agentic follow-up queries.</strong> |
| For multi-hop questions, giving the model a retrieval tool and |
| letting it iterate would outperform the current single-pass |
| design.</li> |
| <li><strong class="text-slate-100">Continuous evaluation.</strong> |
| The eval harness already runs against 16 labelled queries. |
| Running it on every ingest change and diffing the JSON outputs |
| would catch regressions early.</li> |
| </ul> |
| </section> |
|
|
| |
| <section class="mb-12" id="behind"> |
| <h2 class="text-xs uppercase tracking-widest text-slate-500 mb-4">behind the curtain</h2> |
| <p class="text-slate-400 text-sm leading-relaxed"> |
| A live evaluation dashboard with per-query metrics, source-mix |
| breakdowns, latency profile, and run history is available at |
| <a href="/eval" class="text-cyan-300 hover:text-cyan-200 underline decoration-dotted">/eval</a> |
| — password-protected so it doesn't leak eval numbers to casual |
| visitors. Credentials come from the operator's <code class="citation-sample">.env</code>. |
| </p> |
| </section> |
|
|
| <footer class="mt-16 text-xs text-slate-600 text-center"> |
| portfolio demo · no PHI · pgvector + S-PubMedBert + RRF + ms-marco-MiniLM rerank |
| </footer> |
| </main> |
|
|
| <script type="module" src="/static/app.js"></script> |
| <script type="module" src="/static/help.js"></script> |
| </body> |
| </html> |
|
|
