Implement expandable volume breakdown with page chips in search panel

Fix Volume Breakdown query expansion discrepancy in SearchService

Implement hot-reloaded search rules config file for monk prefixes

Fix synonym highlighting in snippet and reader panel, and fix health diagnostics 500 error

Add health diag endpoint for troubleshooting vector search

Implement synonym and related words highlighting in search results and reader panel

Fix write permissions crash on Hugging Face Spaces by configuring paths before downloading and adding a writability check on /data

Fix AI search by appending query expansions instead of replacing the original query

fix: pass task='retrieval.query' to SentenceTransformers encode when using Jina or fine-tuned embedding models to prevent Task missing error

fix: add peft to requirements.txt to support loading fine-tuned LoRA models in SentenceTransformers

feat: support Hugging Face persistent storage at /data and local Docker volumes for persisting SQLite database (search logs/cache) across container rebuilds

feat: FTS5-first + AI button search with split button mode selector

fix: update QWEN_MODEL_ID default to 1.5B in main.py (missed in previous commit)

chore: upgrade Qwen model default from 0.5B to 1.5B Instruct

feat: complete turbovec search backend integration, ignore tvim file in git, and stabilize mobile UI transitions

feat: add Google-style breadcrumb navigation to search results

fix: resolve critical vector embedding bug and refine query transform prompt for prefix word matching

fix: serialize concurrent thread access to in-memory SQLite connection to prevent race conditions and deadlocks

Implement SQLite FTS5 prefix matching to support partial Thai word search and resolve segmentation tokenization issues

Swap default Query-Transform LLM to Qwen2.5-0.5B-Instruct to significantly reduce memory footprint and CPU generation latency on Hugging Face free tier

Optimize PyTorch resource usage: set CPU threads to 1 to prevent health check timeouts and reduce max generation tokens to 45

Optimize backend preloading: preload Qwen & Jina-v5 at startup, cache Qwen in QueryTransformService class attributes, and bypass Ollama connection attempts when not ready

fix: optimize local mode Qdrant restore by replacing server snapshots with pre-packaged SQLite databases

docs: add SEARCH_ARCHITECTURE_V3 spec and include helper scripts for reindexing and RAG query testing

fix: increase Ollama HTTP client timeout to 30.0s to handle model cold starts resiliently

fix: implement query_batch_points and sequential fallback in SearchService, and raise 404 on page not found in pages router

feat: implement Search Architecture V3 (Hybrid search with Qwen query transformation, batch embeddings, batch Qdrant searches, and optimized CPU reranking)

fix: add Cache-Control headers to static files — no-cache for root icons, immutable for hashed assets

fix: serve all root-level static files (icons, favicon.svg, apple-touch-icons)

perf: skip LLM/RAG imports in abbrev lookup, query DB directly

fix: use _default marker_id for non-numeric abbrev ids, skip RAG init

refactor: disable LLM fallback for (ย่อ), fix JSON string array parsing

fix: add อนุ. วิ. ปฏิ. dialogue markers to split formatting

fix: exclude '.' from marker normalization preceding char

fix: merge adjacent <B> headings with <I> footnote refs, fix marker normalization exclusion list

feat: normalize reference extraction and popup UI

feat: enhance footnote extraction and display at bottom of page

feat: implement Reference Lookup API and ReferencePopup integration

feat: migrate jina v5, refine AI popup, add reference popup UI

feat: add sentence-transformers fallback for embedding when Ollama unavailable

fix: cleanup stale Qdrant lock on startup to prevent health alert

fix: extract end markers from raw <B> tags instead of sections (works per-page, not proximity guess)

fix: end markers only show for nearby sections (page proximity filter)

feat: first-page reader layout (centered homage, split volume title) + clean section range markers

feat: use contents table for reader section hierarchy (L0+L1)

feat: use contents table for full section hierarchy in reader (sections array instead of single title)

feat: markdown rendering for AI responses

refactor: clean legacy, sync deps, add embedding cache, split RAGService, split search_service, add tests

Refactor: Make app settings cloud-compatible by removing hardcoded Windows paths

Refactor: Remove remaining hardcoded Windows paths and improve settings flexibility in RAGService