Auto model selection: Haiku classifies question confidence in ~0.6s, then routes high-confidence questions to Haiku (2-5s answer) and complex/debated questions to Sonnet (8-12s). 70% of questions now answer 2-3x faster.
Single Ask button with model dropdown: Auto (smart routing), ⚡ Fast (Haiku), Detailed (Sonnet). Auto is selected by default.
Removed /fast and /medium developer pages. Renamed /deep to /rag for source-grounded evaluation reference.
v0.9.8
Quick Reply + Detailed Answer
Two submit buttons: Quick Reply (Haiku — fast, lower cost) and Detailed Answer (Sonnet — comprehensive, current default). Enter key triggers Detailed Answer.
Confidence spectrum bars are now equal height instead of a rising graph — cleaner visual indicator
v0.9.7
WS1a Confidence Spectrum
Confidence indicator replaced with a 7-level spectrum analyzer: visual red→green signal meter with bars growing taller toward higher confidence
Each bar in the spectrum is individually hoverable with level name and description tooltip
All bars up to the active level are filled; bars beyond are dimmed — like a signal strength meter
Dynamic mode prompt now returns spectrum level labels instead of high/medium/low; legacy values mapped for backward compatibility with Medium and Deep modes
v0.9.6
Clean Mode Routing
Home page now uses Dynamic mode only — no mode selector displayed to beta testers
Developer mode pages available at /deep, /medium, and /fast as direct URLs
Mode badge shown only on non-default developer pages
v0.9.5
Dynamic Mode Fixes + OG Meta Tags
Fixed JSON truncation in Dynamic mode: increased MAX_TOKENS from 2,000 to 4,096 so complete answers with sources and follow-up suggestions render correctly
Dynamic mode now returns a conversation_id and caches context, enabling the follow-up question input box after answers
Scholarly, fatwa, and other non-indexed sources in Dynamic mode answers are now counted as unverified citations — previously showed 0 unverified even when fatwas were cited
Shared links now render dynamic OG meta tags (question + answer preview) for WhatsApp, iMessage, and Twitter card previews instead of the generic site description
Accounting dashboard now shows a By Mode breakdown table: Dynamic vs Medium vs Deep with query count, avg tokens, avg cost, and total cost per mode
Scholarly source tiles now show 🎓 emoji instead of bullet point
Anthropic token counting fix: input_tokens now includes cache_creation and cache_read tokens — was previously reporting only non-cached portion (e.g. 222 tokens instead of actual 56K)
v0.9.4
Dynamic Mode + Citation Verification
Dynamic mode is now the default ask flow: the Answer AI responds first, then Quran and hadith citations are verified against the database
Ask API now accepts a `mode` field (`dynamic`, `medium`, `deep`), keeps `deep_research` as backward-compatible output, and auto-maps old `deep_research=true` requests to `mode='deep'`
Home page now uses a four-button mode selector with Dynamic selected by default, while Fast remains disabled with a Coming soon tooltip
Answer cards now show citation verification totals and color inline citations by verification status, and query logs persist the selected mode
v0.9.3
Scholarly Hadith Numbering
Hadith ingestion now stores up to three numbering systems per narration: `sunnah_com`, `scholarly`, and `in_book` when the CDN provides them
Retriever and vector hadith search now prefer the scholarly number for citations, falling back to the sunnah.com sequence when no scholarly number exists
Dashboard metadata now marks D1-005 complete for the scholarly hadith numbering rollout
v0.9.2
Tafsir-Hadith Cross-Search
Tafsir-referenced hadith cross-search: when tafsir quotes a hadith by collection name, the retriever now finds the matching hadith in the database and adds it to the source list
Primary-source citations improve for tafsir-quoted narrations, enabling direct references such as `Sahih Muslim #1598` instead of indirect tafsir-only attributions
Arabic script collection patterns added: مسلم, بخاري, ترمذی etc. now match in mixed Arabic/English tafsir text
Muslim regex tightened to require 'Sahih Muslim' or Arabic رواہ مسلم — prevents false positives from 'Muslims' (the people) in Islamic text
Adjacent verse scanning (±2 ayahs): cross-search now checks tafsir of nearby verses, catching hadith referenced in consecutive passages
Combined keyword strategy: action verb + topic word (e.g. 'cursed' + 'interest') for cross-translation hadith matching
Stricter include_arabic prompt: only true for Arabic language/recitation questions, not for English questions about Islamic concepts with Arabic names
v0.9.1
Follow-Up Query Mode + Prompt Caching
Answers now return a `conversation_id`, and the home page keeps a follow-up composer under the answer for context-aware follow-up questions
New `/api/v1/ask/follow-up` path skips Search AI and retrieval, reuses cached authenticated source context for 30 minutes, and chains follow-ups on the same conversation
Anthropic answer calls now mark the system prompt and cached source context as ephemeral prompt-cache blocks to reduce repeated follow-up input cost
Query logs now flag follow-up requests separately, with zero search tokens recorded for accounting visibility
v0.9.0
Hadith Citation Priority + Safari Clipboard Fix
Answer AI Rule 10: hadith quoted within tafsir now cited from primary collection (e.g. Sahih Muslim #4093) instead of through the tafsir
Safari clipboard fix: Share button now uses ClipboardItem with Promise for Safari compatibility
Search AI prompt updated to reflect expanded corpus: 7 hadith collections, 5 tafsir works
v0.8.9
Dynamic Retrieval Profiles
Search AI now classifies question type and emits a retrieval profile with preferred tafsir works, Arabic inclusion, and tafsir depth for each query
Standard retrieval now adapts per question: simple questions stay lean, while contested or comparative questions pull broader preferred tafsir coverage
Arabic context gating now follows the retrieval profile in standard mode, while Deep Research remains the unchanged full-source control path
v0.8.8
Token Cost Reduction
Standard mode now uses relevance-ranked context: the top 2 Quran verses and top 2 hadith keep Arabic, while lower-ranked sources switch to English-only formatting
Top-ranked verses now keep full Ibn Kathir commentary, while supporting verses use a lean Ibn Kathir section capped at 3,000 characters after deduplication
Deep Research remains the unconstrained control path with full Arabic and full tafsir for all retrieved sources, and dashboard metadata reflects the amended D1-001 rollout