OrthoPages Physical Evidence

Unlike consumer chatbots that invent non-existent authors or journal papers, OrthoPages enforces strict 1:1 grounding: every cited statement is mapped to a verified physical page index [1], [2] from the retrieved context. The Gemini prompt explicitly forbids hallucinating external bibliography numbers. Clicking any citation badge instantly renders the exact physical textbook page in the viewer for verification.

Our ranking engine applies an Evidence-Based Medicine (EBM) Authority and Recency multiplier. Living evidence synthesis (UpToDate v34.272.0) and specialty practice guidelines (AAOS CPGs, AUCs, OTA) receive a calibrated boost (+4% to +8%) over static textbook chapters. This ensures tie-breaking favors contemporary medical consensus without inverting fundamental clinical relevance.

Stage 1 executes in parallel: SQLite FTS5 BM25 lexical search finds exact medical codes and eponyms, while a 1024-dimensional memory-mapped matrix (voyage-4-large) finds semantic clinical concepts. Their results are merged using weighted Reciprocal Rank Fusion (wRRF). In Stage 2, the top 30 fused candidate pages are scored by the voyage-rerank-3 cross-encoder using hierarchical chapter and section breadcrumbs, and adjacent contiguous pages from the same chapter are stitched together.

The query parser employs strict medical intent discrimination:

• Boolean Triggering: Strictly UPPERCASE AND, OR, and NOT (e.g. "femur AND (IMN OR plate)") trigger Boolean logic trees.
• Conversational Safety: Lowercase and, or, not, and symbols like & or / are treated as natural clinical text. Queries like "Giant Cell Tumor Treatment Non Operative (Conservative & Pharmacologic), Operative" execute smoothly without syntax errors or accidental exclusion.
• Punctuation Normalization: Conversational parentheses (e.g. acronyms like "(CTS)" or qualifiers) are unwrapped into plain text, non-MeSH brackets/braces (e.g. "[1]", "{calcaneus}") are stripped, and single quotes in eponyms (e.g. "Colles' fracture", "Chapman's") are sanitized to protect SQL literals while preserving search precision.

When searching "X vs Y" (e.g. "sinus tarsi vs lateral extensile"), standard vector search frequently tilts heavily toward one procedure if that technique has 10x more pages in textbooks. Our decomposition engine extracts Arm A and Arm B into parallel sub-queries, independently retrieves candidate pools for each arm, and interleaves them before cross-encoder reranking. This guarantees both options have equal representation in the synthesized clinical evidence.

Vignette queries describe demographics, trauma mechanisms, and physical exam findings rather than formal disease names (e.g. "68yo female fell outstretched hand dinner fork deformity"). The engine detects these clinical hallmarks and maps them to canonical surgical entities (Colles fracture, distal radius fracture) without modifying the user's raw prompt, allowing lexical and dense retrieval to find the definitive surgical chapters immediately.

When a query yields zero local candidates or the top rerank score is <0.40 (indicating the topic is novel or not well-represented in static textbooks), Corrective RAG triggers automatically. It queries the NCBI PubMed / MEDLINE API via keyless E-utilities, formats peer-reviewed abstracts as synthetic online sources, and feeds them into the Gemini synthesis engine. If the user has applied a specific book filter or is in offline mode, CRAG fallback is suppressed to respect the user's isolation boundary.

Simply include the word "approach", "exposure", or a specific eponym (e.g. "Hardinge approach", "Thompson approach", "Deltopectoral approach danger zones"). The query triggers our Structured Surgical Directive, instructing Gemini to formulate an Executive Approach Card with: (1) Internervous and intermuscular intervals, (2) Incision bony landmarks, (3) Superficial and deep dissection steps, (4) Specific structures at risk with anatomical safe zones, and (5) Proximal/distal extension methods.

In medical textbooks, a chapter on Giant Cell Tumor of Bone will frequently mention Aneurysmal Bone Cyst under its Differential Diagnosis section. Standard vector search might pull this page when searching for ABC. Our distractor detector scans section titles, headers, and text structure; if the queried condition appears only as a differential mimic on an alternative pathology page, the document is tagged with is_differential_distractor="true", instructing the LLM never to attribute that page's prognosis or recurrence rate to the primary query.

Yes. Select any book from the "Book Filter" dropdown in the Search Deck or click a book title in the catalog. The engine restricts retrieval exclusively to that work's physical pages and injects a Book Isolation Directive into the synthesis prompt, ensuring 0% foreign source leakage.

We pre-extracted 109,685 clinical figures, radiographs, diagrams, and surgical drawings from the master library and embedded them using voyage-multimodal-3.5. When you click the "Diagrams & Algorithms" tab in the workspace, the system executes visual vector search, displaying high-resolution anatomical drawings, surgical positioning setups, and classification schematics.

Click 🎴 Anki Studio in the header or the "Send to Anki" button beneath any AI synthesis. The system automatically connects to your local Anki instance via Anki-Connect (port 8765), converts clinical facts into cloze-deletion and basic Q&A flashcard notes, and pushes them directly into your designated orthopaedic deck.

Quiz Arena generates multiple-choice clinical questions calibrated to residency entrance exams and the orthopaedic board curriculum. Each question features a patient vignette, four distractors with a single correct key, a detailed clinical explanation, and clickable citations mapping back to the authoritative textbook pages.

Yes. Select Force Offline (BM25 Lexical) from the header dropdown. The system operates entirely on your local machine using SQLite FTS5 BM25 search across all 107,373 registered physical pages, allowing full book reading and citation verification even in hospital basements or remote locations with zero internet connectivity.

In the Settings Modal under "Security & Access Control", you can enable Master Passcode Protection. When enabled, the backend enforces constant-time HMAC comparison on all API endpoints and PDF requests. Unauthorized browsers see an encrypted lockscreen overlay until the authorized passcode is provided.

Click ⚙️ Settings in the header and navigate to the "API Keys & Engine" tab. You can inspect active keys, view load-balancing telemetry across the Voyage pool, and enter your personal Google Gemini or Xiaomi MiMo API keys with one-click validation.

In clinical surgery and academic board examinations, vague chapter references (e.g., "see Chapter 4") are unacceptable. Surgical safety demands absolute physical provenance. OrthoPages maps every physical sheet of paper to its exact printed book page number, allowing immediate cross-verification in print libraries and physical exam reading lists.