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

1. Clean legacy: removed legacy/, scratch/, test_clean/, __pycache__/,
artifacts/, root-level junk (gitignored)
2. Sync requirements: deleted stale root requirements.txt, added pytest
3. Embedding LRU cache: OrderedDict 256 entries in RAGService._get_embedding
4. RAGService refactor: __init__ -> _init_qdrant, _init_qdrant_local, _init_qdrant_server
5. Split search_service.py: extracted pali_utils.py (516->249 lines)
6. Test infra: 3 test files (260 lines) with pytest

.gitignore

@@ -49,8 +49,8 @@ _inspect_nb.py
 _test_extract.py
 _test_pdf.py
 
-# Test scripts
-test_*.py
+# Test scripts (root-level only)
+/test_*.py
 
 # Backup files
 *.bak.py

webapp/tipitaka-api/app/services/pali_utils.py

@@ -0,0 +1,281 @@
+"""
+Pali/Thai text utilities for the Tipitaka RAG system.
+
+Contains character confusion maps, spelling correctors, and curated
+vocabulary lists used by the search pipeline.
+
+All module-level constants are computed once at import time (zero runtime cost).
+"""
+
+from difflib import SequenceMatcher
+from typing import List
+
+# ── Optional: PyThaiNLP for general Thai spell-checking ───────────
+try:
+    from pythainlp.spell import correct as pythai_spell_correct
+    from pythainlp.tokenize import word_tokenize as pythai_tokenize
+    HAS_PYTHAINLP = True
+except ImportError:
+    HAS_PYTHAINLP = False
+
+# ── Thai Pali Character Confusion Map ─────────────────────────────
+# These are character-level substitution errors commonly made when
+# typing Pali terms in Thai script (คนไทยมักพิมพ์ส/ศ/ษ สลับกัน ฯลฯ)
+THAI_PALI_CONFUSIONS = {
+    # ส group — all map to /s/ sound
+    "ส": ["ศ", "ษ"],
+    "ศ": ["ส", "ษ"],
+    "ษ": ["ส", "ศ"],
+    # ฏ group — retroflex / dental stops
+    "ฏ": ["ต", "ฐ"],
+    "ฐ": ["ต", "ฏ", "ถ"],
+    "ถ": ["ต", "ฐ", "ฑ"],
+    "ฑ": ["ต", "ฐ", "ถ", "ฒ"],
+    "ฒ": ["ด", "ฑ", "ท"],
+    # ณ / น
+    "ณ": ["น"],
+    "น": ["ณ"],
+    # ภ / พ
+    "ภ": ["พ"],
+    "พ": ["ภ"],
+    # ฬ / ล
+    "ฬ": ["ล"],
+    "ล": ["ฬ"],
+}
+
+
+def _generate_pali_variants(word: str) -> set[str]:
+    """Generate common misspellings for a Pali term via single-character substitutions.
+
+    For each character in `word`, if it belongs to a confusion group, replace it
+    with each alternative to produce a likely misspelling. Only single-character
+    changes are generated (not combinatorial) to keep the map lean.
+    """
+    variants: set[str] = set()
+    for i, ch in enumerate(word):
+        if ch in THAI_PALI_CONFUSIONS:
+            for alt in THAI_PALI_CONFUSIONS[ch]:
+                variant = word[:i] + alt + word[i + 1:]
+                if variant != word:
+                    variants.add(variant)
+    return variants
+
+
+def _pythai_autocorrect(q: str) -> str:
+    """Use PyThaiNLP to correct general Thai spelling errors (not Pali-specific).
+
+    Falls back to the original query if PyThaiNLP is unavailable or raises.
+    Only applies corrections when at least one token changes.
+    """
+    if not HAS_PYTHAINLP or not q:
+        return q
+    try:
+        # Tokenize then spell-correct each token independently
+        tokens = pythai_tokenize(q)
+        corrected_tokens = []
+        changed = False
+        for tok in tokens:
+            if len(tok) <= 1:  # Skip single-char tokens (spaces, punctuation)
+                corrected_tokens.append(tok)
+                continue
+            suggestions = pythai_spell_correct(tok)
+            if suggestions and suggestions[0] != tok:
+                corrected_tokens.append(suggestions[0])
+                changed = True
+            else:
+                corrected_tokens.append(tok)
+        if changed:
+            return "".join(corrected_tokens)
+    except Exception:
+        pass
+    return q
+
+
+def _similar_enough(original: str, corrected: str) -> bool:
+    """Check if PyThaiNLP correction is reasonably similar to the original.
+
+    PyThaiNLP uses a general Thai dictionary and does NOT understand Pali
+    specialized vocabulary. It can aggressively "correct" Pali technical terms
+    into unrelated Thai words.
+
+    Examples of destructive corrections we reject:
+      ปุพเพนิวาสานุสสติญาณ (20 chars) → ปพเนวสสตญ (9 chars)  — dropped 55% of text
+      สติปัฏฐาน (9 chars) → correct variant (still ~9 chars)   — accepted
+
+    We use TWO independent guards:
+      1. Length ratio: corrected must be ≥45% of original length (rejects dropped chars)
+      2. SequenceMatcher ratio: ≥0.65 character overlap (rejects garbled text)
+
+    Both must pass for the correction to be accepted.
+    """
+    if original == corrected or not original or not corrected:
+        return True
+    # Guard 1: Length ratio — if corrected is drastically shorter, it's a lossy "correction"
+    len_ratio = min(len(original), len(corrected)) / max(len(original), len(corrected))
+    if len_ratio < 0.45:
+        return False
+    # Guard 2: Edit-distance ratio — actual character overlap
+    # SequenceMatcher.ratio() = 2*M / (len(a) + len(b)) where M = matching chars
+    if SequenceMatcher(None, original, corrected, autojunk=False).ratio() < 0.65:
+        return False
+    return True
+
+
+# ── Pali/Dhamma Auto-correction (full list from v2.1) ─────────────
+PALI_CORRECTIONS: dict[str, str] = {
+    # อริยสัจ
+    "อริยะสัจ": "อริยสัจ", "อริยะสัจ4": "อริยสัจ", "อริยสัจ4": "อริยสัจ ๔",
+    "สี่อริยสัจ": "อริยสัจ", "4อริยสัจ": "อริยสัจ",
+    # นิพพาน
+    "นิพพาณ": "นิพพาน", "นิพาน": "นิพพาน", "นิรพาน": "นิพพาน", "นิบาน": "นิพพาน",
+    # กรรม / ธรรม
+    "กรม": "กรรม", "ธรม": "ธรรม", "ธรรมะ": "ธรรม",
+    # สังสารวัฏ
+    "สังสาวัฎ": "สังสารวัฏ", "วัฎสังสาร": "สังสารวัฏ", "สังสารวัฎ": "สังสารวัฏ",
+    # ไตรลักษณ์
+    "ไตรลักณ์": "ไตรลักษณ์", "ไตรลักษน์": "ไตรลักษณ์", "ไตรลัษณ์": "ไตรลักษณ์",
+    "3ลักษณะ": "ไตรลักษณ์", "สามลักษณะ": "ไตรลักษณ์",
+    # อนิจจัง / ทุกขัง / อนัตตา
+    "อนิตย์": "อนิจจัง", "อนิตยา": "อนิจจัง",
+    "ทุกขะ": "ทุกข์",
+    "อนัตตะ": "อนัตตา", "อนาตา": "อนัตตา",
+    # สติปัฏฐาน
+    "สติปฏฐาน": "สติปัฏฐาน", "สติปัฐฐาน": "สติปัฏฐาน", "สติปัฏฐาน4": "สติปัฏฐาน ๔",
+    # มรรค
+    "มัก": "มรรค", "มาก": "มรรค",
+    "อัฎฐมรรค": "อัฏฐมรรค", "มรรค8": "อัฏฐมรรค",
+    "8มรรค": "อัฏฐมรรค", "มรรคแปด": "อัฏฐมรรค",
+    # โพชฌงค์
+    "โพชงค์": "โพชฌงค์", "โพชฌงค7": "โพชฌงค์",
+    # ปฏิจจสมุปบาท
+    "ปฏิจสมุปบาท": "ปฏิจจสมุปบาท", "ปฎิจจสมุปบาท": "ปฏิจจสมุปบาท",
+    "ปัจจยาการ": "ปฏิจจสมุปบาท",
+    # กิเลส / ขันธ์
+    "กิเลษ": "กิเลส", "กิเล็ส": "กิเลส",
+    "ขัน5": "ขันธ์ ๕", "5ขันธ์": "ขันธ์ ๕", "ขันธ์5": "ขันธ์ ๕", "ขันห์": "ขันธ์",
+    # โพธิ / วินัย / อภิธรรม
+    "โพธิ์": "โพธิ", "โพธิ์สัตว์": "โพธิสัตว์", "พระโพธิ์สัตว์": "โพธิสัตว์",
+    "วินัยปิฏก": "วินัยปิฎก", "วินัยปิฎฎก": "วินัยปิฎก",
+    "อภิธรม": "อภิธรรม", "อภิธัมม์": "อภิธรรม",
+    # สุตตันตปิฎก
+    "สุตันตปิฎก": "สุตตันตปิฎก", "สูตรปิฎก": "สุตตันตปิฎก",
+    # สมาธิ / วิปัสสนา / สมถะ
+    "สมาธิ์": "สมาธิ",
+    "วิปัสนา": "วิปัสสนา", "วิปาสสนา": "วิปัสสนา", "วิปสสนา": "วิปัสสนา",
+    "สมถ": "สมถะ", "สมธะ": "สมถะ",
+    # ปรินิพพาน
+    "ปรินิพาน": "ปรินิพพาน", "ปรินิพพาณ": "ปรินิพพาน",
+    # มรณสติ
+    "มรนสติ": "มรณสติ",
+    # มหาบุรุษ
+    "มหาบุรษ": "มหาบุรุษ",
+}
+
+
+def to_thai_digits(s: str) -> str:
+    return s.translate(str.maketrans("0123456789", "๐๑๒๓๔๕๖๗๘๙"))
+
+
+# ── Curated popular terms (fallback when search_log is empty) ──────
+# Source: https://84000.org/tipitaka/dic/ (topic headings, stripped of varga numbers)
+POPULAR_TERMS: List[str] = [
+    # ── Core doctrines ──────────────────────────────────────────
+    "อริยสัจ", "นิพพาน", "ไตรลักษณ์", "อนิจจัง", "ทุกข์", "อนัตตา",
+    "สติปัฏฐาน", "อัฏฐมรรค", "โพชฌงค์", "ปฏิจจสมุปบาท",
+    "สมาธิ", "วิปัสสนา", "สมถะ", "ศีล", "ปัญญา",
+    "กรรม", "วิบาก", "กิเลส", "ตัณหา", "อวิชชา",
+    "โพธิสัตว์", "ปรินิพพาน", "สังสารวัฏ", "มรณสติ",
+    "เมตตา", "กรุณา", "มุทิตา", "อุเบกขา",
+    "อภิธรรม", "วินัยปิฎ��", "สุตตันตปิฎก",
+    "ขันธ์", "อายตนะ", "ธาตุ", "อินทรีย์",
+    "โพธิปักขิยธรรม", "สัมมาสมาธิ", "สัมมาสติ",
+    "บารมี", "เจตสิก", "จิต", "รูป", "นาม",
+    # ── From 84000 dictionary ───────────────────────────────────
+    "กถาวัตถุ", "กรรมกิเลส", "กรรมฐาน", "กสิณ",
+    "กัลยาณมิตตตา", "กัลยาณมิตรธรรม", "กาม", "กามคุณ",
+    "กามโภคี", "กาลามสูตรกังขานิยฐาน", "กิจ", "กิจในอริยสัจจ์",
+    "กุลจิรัฏฐิติธรรม", "กุศลกรรมบถ", "กุศลมูล", "กุศลวิตก",
+    "คารวะ", "ฆราวาสธรรม", "จรณะ", "จริต", "จักขุ",
+    "จักร", "จักรวรรดิวัตร", "ฌาน", "ฌาน 2 ประเภท", "ญาณ",
+    "ถูปารหบุคคล", "ทวาร", "ทศพลญาณ", "ทักขิณาวิสุทธิ",
+    "ทาน", "ทิฏฐธัมมิกัตถสังวัตตนิกธรรม", "ทิฏฐิ", "ทิศ",
+    "ที่สุด", "ทุกขตา", "ทุจริต", "ธรรม", "ธรรมขันธ์",
+    "ธรรมคุณ", "ธรรมคุ้มครองโลก", "ธรรมทำให้งาม", "ธรรมนิยาม",
+    "ธรรมมีอุปการะมาก", "ธรรมสมาทาน", "ธรรมสมาธิ",
+    "ธรรมสวนานิสงส์", "ธรรมเทสกธรรม", "ธาตุกัมมัฏฐาน",
+    "ธุดงค์", "ธุระ", "นวกภิกขุธรรม", "นวังคสัตถุสาสน์",
+    "นาถกรณธรรม", "นิมิต", "นิยาม", "นิวรณ์", "นิโรธ",
+    "บริษัท", "บัญญัติ 2", "บุคคล", "บุคคลหาได้ยาก",
+    "บุญกิริยาวัตถุ", "บุตร", "บุพนิมิตแห่งมรรค", "บูชา",
+    "ปฏิปทา", "ปฏิสันถาร", "ปฏิสัมภิทา", "ปธาน",
+    "ปปัญจะ, ปปัญจธรรม", "ปรมัตถธรรม", "ประมาณ", "ปริญญา",
+    "ปริเยสนา", "ปัจจัย", "ปัจจัยให้เกิดสัมมาทิฏฐิ",
+    "ปัพพชิตอภิณหปัจจเวกขณ์", "ปาฏิหาริย์", "ปาปณิกธรรม",
+    "ปาพจน์", "ปาริสุทธิศีล", "ปิยรูป สาตรูป", "ปีติ",
+    "ผล", "พร", "พรหมวิหาร", "พละ",
+    "พละ 5 ของพระมหากษัตริย์", "พหูสูตมีองค์", "พุทธคุณ",
+    "พุทธจริยา", "พุทธโอวาท", "ภพ", "ภรรยา",
+    "ภัพพตาธรรม", "ภาวนา", "ภูมิ", "มงคล", "มรรค",
+    "มรรคมีองค์", "มละ", "มหาปเทส",
+    "มหาปเทส 4 เฉพาะในทางพระวินัย", "มหาภูต", "มัจฉริยะ",
+    "มานะ", "มาร", "มิจฉัตตะ", "มิตรปฏิรูปก์",
+    "รัตนตรัย", "ราชธรรม", "ฤทธิ์", "ลักษณะตัดสินธรรมวินัย",
+    "ลัทธินอกพระพุทธศาสนา", "ลีลาการสอน", "วณิชชา", "วรรณะ",
+    "วัฏฏะ", "วัฒนมุข", "วัตถุประสงค์ในการบัญญัติวินัย",
+    "วิชชา", "วิญญาณ", "วิญญาณฐิติ",
+    "วิธีปฏิบัติต่อทุกข์-สุข", "วิบัติ", "วิปัลลาส",
+    "วิปัสสนาญาณ", "วิปัสสนูปกิเลส", "วิมุตติ", "วิรัติ",
+    "วิสุทธิ", "วิเวก", "วิโมกข์", "วุฒิ", "ศรัทธา",
+    "ศีล 8 ทั้งอาชีวะ", "สกทาคามี", "สมชีวิธรรม", "สมบัติ",
+    "สมาธิภาวนา", "สมาบัติ", "สรณะ", "สวรรค์",
+    "สังขตลักษณะ", "สังขาร", "สังคหวัตถุ",
+    "สังคห���ัตถุของผู้ครองแผ่นดิน", "สังคหะ", "สังฆคุณ",
+    "สังวร", "สังเวชนียสถาน", "สังโยชน์", "สัจจะ", "สัญญา",
+    "สัตตาวาส", "สัทธรรม", "สันโดษ", "สัปปายะ",
+    "สัปปุริสทาน", "สัปปุริสธรรม", "สัปปุริสบัญญัติ",
+    "สัมปชัญญะ", "สัมปทา", "สัมปรายิกัตถสังวัตตนิกธรรม",
+    "สัมผัส", "สัมมัตตะ", "สารณียธรรม", "สาสน์", "สิกขา",
+    "สุข", "สุขของคฤหัสถ์", "สุจริต", "สุทธาวาส", "สุทธิ",
+    "สุหทมิตร", "อกุศลกรรมบถ", "อกุศลมูล", "อกุศลวิตก",
+    "อคติ", "อธิปไตย", "อธิษฐาน", "อนันตริยกรรม",
+    "อนาคามี", "อนุตตริยะ", "อนุบุพพวิหาร", "อนุปุพพิกถา",
+    "อนุสติ", "อนุสัย", "อบาย", "อบายมุข",
+    "อปริหานิยธรรม", "อปัณณกปฏิปทา", "อปัสเสนะ",
+    "อภิญญา", "อภิฐาน", "อภิณหปัจจเวกขณ์", "อภิสังขาร",
+    "อรหันต์", "อริยทรัพย์", "อริยบุคคล", "อริยวงศ์",
+    "อริยวัฑฒิ", "อริยสัจจ์", "อรูป", "อสังขตลักษณะ",
+    "อสุภะ", "อัคคิ", "อัตถะ", "อันตคาหิกทิฏฐิ",
+    "อันตรายของภิกษุสามเณรผู้บวชใหม่", "อัปปมาทะ",
+    "อาการที่พระพุทธเจ้าทรงสั่งสอน", "อาจารย์",
+    "อานาปานสติ 16 ฐาน", "อายตนะภายนอก", "อายตนะภายใน",
+    "อายุสสธรรม", "อาวาสิกธรรม", "อาสวะ", "อาหาร",
+    "อิทธิบาท", "อุบาสกธรรม", "อุปกิเลส", "อุปัญญาตธรรม",
+    "อุปาทาน", "อุปาทารูป", "เจดีย์", "เจตนา", "เถรธรรม",
+    "เทพ", "เทวทูต", "เทศนา", "เบญจธรรม", "เมถุนสังโยค",
+    "เวทนา", "เวปุลลธรรม", "เวปุลละ", "เวสารัชชกรณธรรม",
+    "เวสารัชชะ", "โกศล", "โภควิภาค", "โภคอาทิยะ",
+    "โยคะ", "โยนิ", "โยนิโสมนสิการ", "โลก", "โลกธรรม",
+    "โลกุตตรธรรม", "โสดาบัน", "โสตาปัตติยังคะ", "โอฆะ",
+    "ไตรปิฎก",
+]
+
+
+def _build_pali_corrections() -> dict[str, str]:
+    """Extend PALI_CORRECTIONS with auto-generated variants from all POPULAR_TERMS.
+
+    For every term in POPULAR_TERMS, this generates likely misspellings
+    using character confusion groups (ส↔ศ↔ษ, ฏ↔ต↔ฐ, ณ↔น, ภ↔พ, ฬ↔ล, etc.)
+    and adds them to the corrections dict.
+
+    This is called once at module load time so there is zero runtime cost.
+    """
+    corrections = dict(PALI_CORRECTIONS)
+    for term in POPULAR_TERMS:
+        for variant in _generate_pali_variants(term):
+            if variant not in corrections:
+                corrections[variant] = term
+    return corrections
+
+
+# Build the full corrections dict at load time (hand-curated + auto-generated)
+PALI_CORRECTIONS_FULL = _build_pali_corrections()

webapp/tipitaka-api/app/services/rag_service.py

@@ -1,5 +1,6 @@
 import logging
 from pathlib import Path
+from collections import OrderedDict
 import qdrant_client
 from qdrant_client.http import models as qmodels
 import httpx
@@ -48,91 +49,111 @@ class RAGService:
         self.model = None  # Flag: None = not verified, 'ready' = OK
         self.reranker = None
         self.actual_chunks_col = "tipitaka_chunks" # Default name
-        
+        self._embed_cache = OrderedDict()     # LRU cache: text → embedding
+        self._embed_cache_max = 256           # max cached entries
+        self.client = None
+
+        # Delegate heavy Qdrant init to a dedicated method
+        self._init_qdrant()
+
+        # Pre-load models (Embedding + Reranker) — safe to fail gracefully
+        self._load_model()
+        self._load_reranker()
+
+    def _init_qdrant(self):
+        """Heavy initialization: detect mode, extract snapshots, create Qdrant client."""
+        settings = get_settings()
+        qdrant_url = getattr(settings, "QDRANT_URL", None)
+
+        # Auto-detect Qdrant server if not configured
+        if not qdrant_url:
+            try:
+                with httpx.Client() as client:
+                    response = client.get("http://localhost:6333/healthz", timeout=1.0)
+                    if response.status_code == 200:
+                        qdrant_url = "http://localhost:6333"
+                        logger.info(f"Auto-detected running Qdrant Server at {qdrant_url}")
+            except Exception:
+                pass
+
+        is_local = not bool(qdrant_url)
+
         try:
-            # 1. Determine Mode (Auto-detect Server vs Local)
-            qdrant_url = getattr(settings, "QDRANT_URL", None)
-            
-            if not qdrant_url:
-                try:
-                    with httpx.Client() as client:
-                        response = client.get("http://localhost:6333/healthz", timeout=1.0)
-                        if response.status_code == 200:
-                            qdrant_url = "http://localhost:6333"
-                            logger.info(f"Auto-detected running Qdrant Server at {qdrant_url}")
-                except Exception:
-                    pass
-
-            is_local = not bool(qdrant_url)
-            
             if is_local:
-                logger.info(f"Initializing Qdrant in Local Mode at {self.qdrant_path}")
-                storage_path = Path(self.qdrant_path)
-                storage_path.mkdir(parents=True, exist_ok=True)
-                
-                lock_file = storage_path / ".lock"
-                if lock_file.exists():
-                    try:
-                        logger.warning(f"Removing stale Qdrant lock file: {lock_file}")
-                        lock_file.unlink()
-                    except Exception as e:
-                        logger.error(f"Failed to remove lock file: {e}")
-
-                for col_name in self.collections:
-                    col_dir = storage_path / "collections" / col_name
-                    if not col_dir.exists():
-                        snap_path = self.snapshot_dir / f"{col_name}.snapshot"
-                        if snap_path.exists():
-                            logger.info(f"Restoring '{col_name}' via manual extraction...")
-                            self._extract_snapshot(col_name, snap_path)
-                
-                self.client = qdrant_client.QdrantClient(path=self.qdrant_path)
-                logger.info("Qdrant Local Client initialized.")
-            
+                self._init_qdrant_local()
             else:
-                logger.info(f"Connecting to Qdrant Server at {qdrant_url}")
-                self.client = qdrant_client.QdrantClient(url=qdrant_url)
-                
-                try:
-                    all_cols = [c.name for c in self.client.get_collections().collections]
-                except Exception as e:
-                    logger.error(f"Failed to list collections: {e}")
-                    all_cols = []
-
-                for col_name in self.collections:
-                    actual_col = None
-                    if col_name in all_cols:
-                        actual_col = col_name
-                    else:
-                        matches = [c for c in all_cols if c.startswith(f"{col_name}_") or c.startswith(col_name)]
-                        if matches: actual_col = matches[0]
-
-                    if actual_col:
-                        if col_name == "tipitaka_chunks": self.actual_chunks_col = actual_col
-                    else:
-                        snap_filename = f"{col_name}.snapshot"
-                        target_snap = ALLOWED_SNAP_ROOT / snap_filename
-                        if not target_snap.exists(): target_snap = self.snapshot_dir / snap_filename
-                            
-                        if target_snap.exists():
-                            import os
-                            logger.info(f"Restoring server collection '{col_name}' from {target_snap}...")
-                            abs_snap_path = os.path.abspath(target_snap).replace("\\", "/")
-                            if not abs_snap_path.startswith("/"): abs_snap_path = "/" + abs_snap_path
-                            try:
-                                self.client.recover_snapshot(col_name, location=f"file://{abs_snap_path}")
-                                if col_name == "tipitaka_chunks": self.actual_chunks_col = col_name
-                            except Exception as e:
-                                logger.error(f"Failed to restore: {e}")
-
-            # 3. Pre-load Models (Embedding + Reranker)
-            self._load_model()
-            self._load_reranker()
-
+                self._init_qdrant_server(qdrant_url)
         except Exception as e:
             logger.error(f"RAG Initialization Error: {e}")
             self.client = None
 
+    def _init_qdrant_local(self):
+        """Initialize Qdrant in Local Mode — extract snapshots if needed."""
+        logger.info(f"Initializing Qdrant in Local Mode at {self.qdrant_path}")
+        storage_path = Path(self.qdrant_path)
+        storage_path.mkdir(parents=True, exist_ok=True)
+
+        lock_file = storage_path / ".lock"
+        if lock_file.exists():
+            try:
+                logger.warning(f"Removing stale Qdrant lock file: {lock_file}")
+                lock_file.unlink()
+            except Exception as e:
+                logger.error(f"Failed to remove lock file: {e}")
+
+        for col_name in self.collections:
+            col_dir = storage_path / "collections" / col_name
+            if not col_dir.exists():
+                snap_path = self.snapshot_dir / f"{col_name}.snapshot"
+                if snap_path.exists():
+                    logger.info(f"Restoring '{col_name}' via manual extraction...")
+                    self._extract_snapshot(col_name, snap_path)
+
+        self.client = qdrant_client.QdrantClient(path=self.qdrant_path)
+        logger.info("Qdrant Local Client initialized.")
+
+    def _init_qdrant_server(self, qdrant_url: str):
+        """Initialize Qdrant in Server Mode — connect and restore snapshots if needed."""
+        logger.info(f"Connecting to Qdrant Server at {qdrant_url}")
+        self.client = qdrant_client.QdrantClient(url=qdrant_url)
+
+        try:
+            all_cols = [c.name for c in self.client.get_collections().collections]
+        except Exception as e:
+            logger.error(f"Failed to list collections: {e}")
+            all_cols = []
+
+        for col_name in self.collections:
+            actual_col = None
+            if col_name in all_cols:
+                actual_col = col_name
+            else:
+                matches = [c for c in all_cols if c.startswith(f"{col_name}_") or c.startswith(col_name)]
+                if matches:
+                    actual_col = matches[0]
+
+            if actual_col:
+                if col_name == "tipitaka_chunks":
+                    self.actual_chunks_col = actual_col
+            else:
+                snap_filename = f"{col_name}.snapshot"
+                target_snap = ALLOWED_SNAP_ROOT / snap_filename
+                if not target_snap.exists():
+                    target_snap = self.snapshot_dir / snap_filename
+
+                if target_snap.exists():
+                    import os
+                    logger.info(f"Restoring server collection '{col_name}' from {target_snap}...")
+                    abs_snap_path = os.path.abspath(target_snap).replace("\\", "/")
+                    if not abs_snap_path.startswith("/"):
+                        abs_snap_path = "/" + abs_snap_path
+                    try:
+                        self.client.recover_snapshot(col_name, location=f"file://{abs_snap_path}")
+                        if col_name == "tipitaka_chunks":
+                            self.actual_chunks_col = col_name
+                    except Exception as e:
+                        logger.error(f"Failed to restore: {e}")
+
     def _load_model(self):
         """Verify Ollama embedding service is accessible."""
         if self.model is None:
@@ -160,14 +181,26 @@ class RAGService:
                 self.reranker = "error"
 
     def _get_embedding(self, text: str) -> list:
-        """Get embedding vector from Ollama API."""
+        """Get embedding vector from Ollama API, with LRU cache."""
+        # Check LRU cache first
+        if text in self._embed_cache:
+            self._embed_cache.move_to_end(text)  # Mark as recently used
+            return self._embed_cache[text]
+
         response = httpx.post(
             f"{OLLAMA_URL}/api/embed",
             json={"model": EMBED_MODEL, "input": text},
             timeout=30
         )
         response.raise_for_status()
-        return response.json()["embeddings"][0]
+        embedding = response.json()["embeddings"][0]
+
+        # Store in LRU cache
+        self._embed_cache[text] = embedding
+        if len(self._embed_cache) > self._embed_cache_max:
+            self._embed_cache.popitem(last=False)  # Remove oldest (LRU)
+
+        return embedding
 
     async def query(self, text: str, n_results: int = 10, threshold: float = 0.2) -> str:
         """

webapp/tipitaka-api/app/services/search_service.py

@@ -1,279 +1,15 @@
 import time
 import re
-from difflib import SequenceMatcher
 from typing import List
 from app.database.sqlite_db import SQLiteDB
 from app.schemas import SearchResponse, SearchResultItem
-
-# ── Optional: PyThaiNLP for general Thai spell-checking ───────────
-try:
-    from pythainlp.spell import correct as pythai_spell_correct
-    from pythainlp.tokenize import word_tokenize as pythai_tokenize
-    HAS_PYTHAINLP = True
-except ImportError:
-    HAS_PYTHAINLP = False
-
-# ── Thai Pali Character Confusion Map ─────────────────────────────
-# These are character-level substitution errors commonly made when
-# typing Pali terms in Thai script (คนไทยมักพิมพ์ส/ศ/ษ สลับกัน ฯลฯ)
-THAI_PALI_CONFUSIONS = {
-    # ส group — all map to /s/ sound
-    "ส": ["ศ", "ษ"],
-    "ศ": ["ส", "ษ"],
-    "ษ": ["ส", "ศ"],
-    # ฏ group — retroflex / dental stops
-    "ฏ": ["ต", "ฐ"],
-    "ฐ": ["ต", "ฏ", "ถ"],
-    "ถ": ["ต", "ฐ", "ฑ"],
-    "ฑ": ["ต", "ฐ", "ถ", "ฒ"],
-    "ฒ": ["ด", "ฑ", "ท"],
-    # ณ / น
-    "ณ": ["น"],
-    "น": ["ณ"],
-    # ภ / พ
-    "ภ": ["พ"],
-    "พ": ["ภ"],
-    # ฬ / ล
-    "ฬ": ["ล"],
-    "ล": ["ฬ"],
-}
-
-
-def _generate_pali_variants(word: str) -> set[str]:
-    """Generate common misspellings for a Pali term via single-character substitutions.
-
-    For each character in `word`, if it belongs to a confusion group, replace it
-    with each alternative to produce a likely misspelling. Only single-character
-    changes are generated (not combinatorial) to keep the map lean.
-    """
-    variants: set[str] = set()
-    for i, ch in enumerate(word):
-        if ch in THAI_PALI_CONFUSIONS:
-            for alt in THAI_PALI_CONFUSIONS[ch]:
-                variant = word[:i] + alt + word[i + 1:]
-                if variant != word:
-                    variants.add(variant)
-    return variants
-
-
-def _pythai_autocorrect(q: str) -> str:
-    """Use PyThaiNLP to correct general Thai spelling errors (not Pali-specific).
-
-    Falls back to the original query if PyThaiNLP is unavailable or raises.
-    Only applies corrections when at least one token changes.
-    """
-    if not HAS_PYTHAINLP or not q:
-        return q
-    try:
-        # Tokenize then spell-correct each token independently
-        tokens = pythai_tokenize(q)
-        corrected_tokens = []
-        changed = False
-        for tok in tokens:
-            if len(tok) <= 1:  # Skip single-char tokens (spaces, punctuation)
-                corrected_tokens.append(tok)
-                continue
-            suggestions = pythai_spell_correct(tok)
-            if suggestions and suggestions[0] != tok:
-                corrected_tokens.append(suggestions[0])
-                changed = True
-            else:
-                corrected_tokens.append(tok)
-        if changed:
-            return "".join(corrected_tokens)
-    except Exception:
-        pass
-    return q
-
-
-def _similar_enough(original: str, corrected: str) -> bool:
-    """Check if PyThaiNLP correction is reasonably similar to the original.
-
-    PyThaiNLP uses a general Thai dictionary and does NOT understand Pali
-    specialized vocabulary. It can aggressively "correct" Pali technical terms
-    into unrelated Thai words.
-
-    Examples of destructive corrections we reject:
-      ปุพเพนิวาสานุสสติญาณ (20 chars) → ปพเนวสสตญ (9 chars)  — dropped 55% of text
-      สติปัฏฐาน (9 chars) → correct variant (still ~9 chars)   — accepted
-
-    We use TWO independent guards:
-      1. Length ratio: corrected must be ≥45% of original length (rejects dropped chars)
-      2. SequenceMatcher ratio: ≥0.65 character overlap (rejects garbled text)
-
-    Both must pass for the correction to be accepted.
-    """
-    if original == corrected or not original or not corrected:
-        return True
-    # Guard 1: Length ratio — if corrected is drastically shorter, it's a lossy "correction"
-    len_ratio = min(len(original), len(corrected)) / max(len(original), len(corrected))
-    if len_ratio < 0.45:
-        return False
-    # Guard 2: Edit-distance ratio — actual character overlap
-    # SequenceMatcher.ratio() = 2*M / (len(a) + len(b)) where M = matching chars
-    if SequenceMatcher(None, original, corrected, autojunk=False).ratio() < 0.65:
-        return False
-    return True
-
-
-# ── Pali/Dhamma Auto-correction (full list from v2.1) ─────────────
-PALI_CORRECTIONS: dict[str, str] = {
-    # อริยสัจ
-    "อริยะสัจ": "อริยสัจ", "อริยะสัจ4": "อริยสัจ", "อริยสัจ4": "อริยสัจ ๔",
-    "สี่อริยสัจ": "อริยสัจ", "4อริยสัจ": "อริยสัจ",
-    # นิพพาน
-    "นิพพาณ": "นิพพาน", "นิพาน": "นิพพาน", "นิรพาน": "นิพพาน", "นิบาน": "นิพพาน",
-    # กรรม / ธรรม
-    "กรม": "กรรม", "ธรม": "ธรรม", "ธรรมะ": "ธรรม",
-    # สังสารวัฏ
-    "สังสาวัฎ": "สังสารวัฏ", "วัฎสังสาร": "สังสารวัฏ", "สังสารวัฎ": "สังสารวัฏ",
-    # ไตรลักษณ์
-    "ไตรลักณ์": "ไตรลักษณ์", "ไตรลักษน์": "ไตรลักษณ์", "ไตรลัษณ์": "ไตรลักษณ์",
-    "3ลักษณะ": "ไตรลักษณ์", "สามลักษณะ": "ไตรลักษณ์",
-    # อนิจจัง / ทุกขัง / อนัตตา
-    "อนิตย์": "อนิจจัง", "อนิตยา": "อนิจจัง",
-    "ทุกขะ": "ทุกข์",
-    "อนัตตะ": "อนัตตา", "อนาตา": "อนัตตา",
-    # สติปัฏฐาน
-    "สติปฏฐาน": "สติปัฏฐาน", "สติปัฐฐาน": "สติปัฏฐาน", "สติปัฏฐาน4": "สติปัฏฐาน ๔",
-    # มรรค
-    "มัก": "มรรค", "มาก": "มรรค",
-    "อัฎฐมรรค": "อัฏฐมรรค", "มรรค8": "อัฏฐมรรค",
-    "8มรรค": "อัฏฐมรรค", "มรรคแปด": "อัฏฐมรรค",
-    # โพชฌงค์
-    "โพชงค์": "โพชฌงค์", "โพชฌงค7": "โพชฌงค์",
-    # ปฏิจจสมุปบาท
-    "ปฏิจสมุปบาท": "ปฏิจจสมุปบาท", "ปฎิจจสมุปบาท": "ปฏิจจสมุปบาท",
-    "ปัจจยาการ": "ปฏิจจสมุปบาท",
-    # กิเลส / ขันธ์
-    "กิเลษ": "กิเลส", "กิเล็ส": "กิเลส",
-    "ขัน5": "ขันธ์ ๕", "5ขันธ์": "ขันธ์ ๕", "ขันธ์5": "ขันธ์ ๕", "ขันห์": "ขันธ์",
-    # โพธิ / วินัย / อภิธรรม
-    "โพธิ์": "โพธิ", "โพธิ์สัตว์": "โพธิสัตว์", "พระโพธิ์สัตว์": "โพธิสัตว์",
-    "วินัยปิฏก": "วินัยปิฎก", "วินัยปิฎฎก": "วินัยปิฎก",
-    "อภิธรม": "อภิธรรม", "อภิธัมม์": "อภิธรรม",
-    # สุตตันตปิฎก
-    "สุตันตปิฎก": "สุตตันตปิฎก", "สูตรปิฎก": "สุตตันตปิฎก",
-    # สมาธิ / วิปัสสนา / สมถะ
-    "สมาธิ์": "สมาธิ",
-    "วิปัสนา": "วิปัสสนา", "วิปาสสนา": "วิปัสสนา", "วิปสสนา": "วิปัสสนา",
-    "สมถ": "สมถะ", "สมธะ": "สมถะ",
-    # ปรินิพพาน
-    "ปรินิพาน": "ปรินิพพาน", "ปรินิพพาณ": "ปรินิพพาน",
-    # มรณสติ
-    "มรนสติ": "มรณสติ",
-    # มหาบุรุษ
-    "มหาบุรษ": "มหาบุรุษ",
-}
-
-
-def to_thai_digits(s: str) -> str:
-    return s.translate(str.maketrans("0123456789", "๐๑๒๓๔๕๖๗๘๙"))
-
-
-# ── Curated popular terms (fallback when search_log is empty) ──────
-# Source: https://84000.org/tipitaka/dic/ (topic headings, stripped of varga numbers)
-POPULAR_TERMS: List[str] = [
-    # ── Core doctrines ──────────────────────────────────────────
-    "อริยสัจ", "นิพพาน", "ไตรลักษณ์", "อนิจจัง", "ทุกข์", "อนัตตา",
-    "สติปัฏฐาน", "อัฏฐมรรค", "โพชฌงค์", "ปฏิจจสมุปบาท",
-    "สมาธิ", "วิปัสสนา", "สมถะ", "ศีล", "ปัญญา",
-    "กรรม", "วิบาก", "กิเลส", "ตัณหา", "อวิชชา",
-    "โพธิสัตว์", "ปรินิพพาน", "สังสารวัฏ", "มรณสติ",
-    "เมตตา", "กรุณา", "มุทิตา", "อุเบกขา",
-    "อภิธรรม", "วินัยปิฎก", "สุตตันตปิฎก",
-    "ขันธ์", "อายตนะ", "ธาตุ", "อินทรีย์",
-    "โ���ธิปักขิยธรรม", "สัมมาสมาธิ", "สัมมาสติ",
-    "บารมี", "เจตสิก", "จิต", "รูป", "นาม",
-    # ── From 84000 dictionary ───────────────────────────────────
-    "กถาวัตถุ", "กรรมกิเลส", "กรรมฐาน", "กสิณ",
-    "กัลยาณมิตตตา", "กัลยาณมิตรธรรม", "กาม", "กามคุณ",
-    "กามโภคี", "กาลามสูตรกังขานิยฐาน", "กิจ", "กิจในอริยสัจจ์",
-    "กุลจิรัฏฐิติธรรม", "กุศลกรรมบถ", "กุศลมูล", "กุศลวิตก",
-    "คารวะ", "ฆราวาสธรรม", "จรณะ", "จริต", "จักขุ",
-    "จักร", "จักรวรรดิวัตร", "ฌาน", "ฌาน 2 ประเภท", "ญาณ",
-    "ถูปารหบุคคล", "ทวาร", "ทศพลญาณ", "ทักขิณาวิสุทธิ",
-    "ทาน", "ทิฏฐธัมมิกัตถสังวัตตนิกธรรม", "ทิฏฐิ", "ทิศ",
-    "ที่สุด", "ทุกขตา", "ทุจริต", "ธรรม", "ธรรมขันธ์",
-    "ธรรมคุณ", "ธรรมคุ้มครองโลก", "ธรรมทำให้งาม", "ธรรมนิยาม",
-    "ธรรมมีอุปการะมาก", "ธรรมสมาทาน", "ธรรมสมาธิ",
-    "ธรรมสวนานิสงส์", "ธรรมเทสกธรรม", "ธาตุกัมมัฏฐาน",
-    "ธุดงค์", "ธุระ", "นวกภิกขุธรรม", "นวังคสัตถุสาสน์",
-    "นาถกรณธรรม", "นิมิต", "นิยาม", "นิวรณ์", "นิโรธ",
-    "บริษัท", "บัญญัติ 2", "บุคคล", "บุคคลหาได้ยาก",
-    "บุญกิริยาวัตถุ", "บุตร", "บุพนิมิตแห่งมรรค", "บูชา",
-    "ปฏิปทา", "ปฏิสันถาร", "ปฏิสัมภิทา", "ปธาน",
-    "ปปัญจะ, ปปัญจธรรม", "ปรมัตถธรรม", "ประมาณ", "ปริญญา",
-    "ปริเยสนา", "ปัจจัย", "ปัจจัยให้เกิดสัมมาทิฏฐิ",
-    "ปัพพชิตอภิณหปัจจเวกขณ์", "ปาฏิหาริย์", "ปาปณิกธรรม",
-    "ปาพจน์", "ปาริสุทธิศีล", "ปิยรูป สาตรูป", "ปีติ",
-    "ผล", "พร", "พรหมวิหาร", "พละ",
-    "พละ 5 ของพระมหากษัตริย์", "พหูสูตมีองค์", "พุทธคุณ",
-    "พุทธจริยา", "พุทธโอวาท", "ภพ", "ภรรยา",
-    "ภัพพตาธรรม", "ภาวนา", "ภูมิ", "มงคล", "มรรค",
-    "มรรคมีองค์", "มละ", "มหาปเทส",
-    "มหาปเทส 4 เฉพาะในทางพระวินัย", "มหาภูต", "มัจฉริยะ",
-    "มานะ", "มาร", "มิจฉัตตะ", "มิตรปฏิรูปก์",
-    "รัตนตรัย", "ราชธรรม", "ฤทธิ์", "ลักษณะตัดสินธรรมวินัย",
-    "ลัทธินอกพระพุทธศาสนา", "ลีลาการสอน", "วณิชชา", "วรรณะ",
-    "วัฏฏะ", "วัฒนมุข", "วัตถุประสงค์ในการบัญญัติวินัย",
-    "วิชชา", "วิญญาณ", "วิญญาณฐิติ",
-    "วิธีปฏิบัติต่อทุกข์-สุข", "วิบัติ", "วิปัลลาส",
-    "วิปัสสนาญาณ", "วิปัสสนูปกิเลส", "วิมุตติ", "วิรัติ",
-    "วิสุทธิ", "วิเวก", "วิโมกข์", "วุฒิ", "ศรัทธา",
-    "ศีล 8 ทั้งอาชีวะ", "สกทาคามี", "สมชีวิธรรม", "สมบัติ",
-    "สมาธิภาวนา", "สมาบัติ", "สรณะ", "สวรรค์",
-    "สังขตลักษณะ", "สังขาร", "สังคหวัตถุ",
-    "สังคหวัตถุของผู้ครองแผ่นดิน", "สังคหะ", "สังฆคุณ",
-    "สังวร", "���ังเวชนียสถาน", "สังโยชน์", "สัจจะ", "สัญญา",
-    "สัตตาวาส", "สัทธรรม", "สันโดษ", "สัปปายะ",
-    "สัปปุริสทาน", "สัปปุริสธรรม", "สัปปุริสบัญญัติ",
-    "สัมปชัญญะ", "สัมปทา", "สัมปรายิกัตถสังวัตตนิกธรรม",
-    "สัมผัส", "สัมมัตตะ", "สารณียธรรม", "สาสน์", "สิกขา",
-    "สุข", "สุขของคฤหัสถ์", "สุจริต", "สุทธาวาส", "สุทธิ",
-    "สุหทมิตร", "อกุศลกรรมบถ", "อกุศลมูล", "อกุศลวิตก",
-    "อคติ", "อธิปไตย", "อธิษฐาน", "อนันตริยกรรม",
-    "อนาคามี", "อนุตตริยะ", "อนุบุพพวิหาร", "อนุปุพพิกถา",
-    "อนุสติ", "อนุสัย", "อบาย", "อบายมุข",
-    "อปริหานิยธรรม", "อปัณณกปฏิปทา", "อปัสเสนะ",
-    "อภิญญา", "อภิฐาน", "อภิณหปัจจเวกขณ์", "อภิสังขาร",
-    "อรหันต์", "อริยทรัพย์", "อริยบุคคล", "อริยวงศ์",
-    "อริยวัฑฒิ", "อริยสัจจ์", "อรูป", "อสังขตลักษณะ",
-    "อสุภะ", "อัคคิ", "อัตถะ", "อันตคาหิกทิฏฐิ",
-    "อันตรายของภิกษุสามเณรผู้บวชใหม่", "อัปปมาทะ",
-    "อาการที่พระพุทธเจ้าทรงสั่งสอน", "อาจารย์",
-    "อานาปานสติ 16 ฐาน", "อายตนะภายนอก", "อายตนะภายใน",
-    "อายุสสธรรม", "อาวาสิกธรรม", "อาสวะ", "อาหาร",
-    "อิทธิบาท", "อุบาสกธรรม", "อุปกิเลส", "อุปัญญาตธรรม",
-    "อุปาทาน", "อุปาทารูป", "เจดีย์", "เจตนา", "เถรธรรม",
-    "เทพ", "เทวทูต", "เทศนา", "เบญจธรรม", "เมถุนสังโยค",
-    "เวทนา", "เวปุลลธรรม", "เวปุลละ", "เวสารัชชกรณธรรม",
-    "เวสารัชชะ", "โกศล", "โภควิภาค", "โภคอาทิยะ",
-    "โยคะ", "โยนิ", "โยนิโสมนสิการ", "โลก", "โลกธรรม",
-    "โลกุตตรธรรม", "โสดาบัน", "โสตาปัตติยังคะ", "โอฆะ",
-    "ไตรปิฎก",
-]
-
-
-def _build_pali_corrections() -> dict[str, str]:
-    """Extend PALI_CORRECTIONS with auto-generated variants from all POPULAR_TERMS.
-
-    For every term in POPULAR_TERMS, this generates likely misspellings
-    using character confusion groups (ส↔ศ↔ษ, ฏ↔ต↔ฐ, ณ↔น, ภ↔พ, ฬ↔ล, etc.)
-    and adds them to the corrections dict.
-
-    This is called once at module load time so there is zero runtime cost.
-    """
-    corrections = dict(PALI_CORRECTIONS)
-    for term in POPULAR_TERMS:
-        for variant in _generate_pali_variants(term):
-            if variant not in corrections:
-                corrections[variant] = term
-    return corrections
-
-
-# Build the full corrections dict at load time (hand-curated + auto-generated)
-PALI_CORRECTIONS_FULL = _build_pali_corrections()
+from app.services.pali_utils import (
+    PALI_CORRECTIONS_FULL,
+    POPULAR_TERMS,
+    to_thai_digits,
+    _pythai_autocorrect,
+    _similar_enough,
+)
 
 
 class SearchService:
@@ -350,7 +86,7 @@ class SearchService:
                 continue
             if re.match(r"^[_\-=]{3,}$", ln.strip()):  # separator lines
                 continue
-            ln = re.sub(r'[๐๑๒๓๔๕๖๗๘๙]+-', '', ln)  # remove Thai numeral footnote markers (๑-, ๒-, etc.)
+            ln = re.sub(r'[๐๑๒๓๔๕๖๗๘๙]+-', '', ln)  # remove Thai numeral footnote markers
             ln = re.sub(r"  +", " ", ln).strip()
             if ln:
                 lines.append(ln)
@@ -407,13 +143,10 @@ class SearchService:
         # ── 1. Auto-correct pipeline ─────────────────────────────
         q = query.strip()
 
-        # Stage A: Pali-specific correction (explicit dict lookup + auto-generated variants)
+        # Stage A: Pali-specific correction (explicit dict + auto-generated variants)
         corrected_q = PALI_CORRECTIONS_FULL.get(q, q)
 
         # Stage B: General Thai spelling correction via PyThaiNLP
-        # NOTE: PyThaiNLP does NOT understand Pali — it uses a general Thai
-        # dictionary and can corrupt Pali terms into unrelated words. We guard
-        # with _similar_enough() to reject destructive "corrections".
         if corrected_q == q:
             pythai_q = _pythai_autocorrect(q)
             if pythai_q != q and _similar_enough(q, pythai_q):
@@ -425,7 +158,7 @@ class SearchService:
         with self.db.get_connection() as conn:
             cursor = conn.cursor()
 
-            # ── 2. FTS5 search (pages_fts — unified name) ──────────
+            # ── 2. FTS5 search ──────────────────────────────────────
             fts_query = f'"{thai_q}"' if "*" not in thai_q else thai_q
             try:
                 cursor.execute("""

webapp/tipitaka-api/requirements.txt

@@ -14,3 +14,5 @@ transformers>=4.40.0
 numpy>=1.24.0
 httpx>=0.27.0
 anyio>=4.0.0
+pytest>=8.0
+pytest-asyncio>=0.24

webapp/tipitaka-api/tests/__init__.py

@@ -0,0 +1,3 @@
+"""
+Tipitaka API — test suite.
+"""

webapp/tipitaka-api/tests/conftest.py

@@ -0,0 +1,23 @@
+"""
+Shared test fixtures for Tipitaka API tests.
+"""
+
+import pytest
+from unittest.mock import MagicMock
+
+
+@pytest.fixture
+def mock_db():
+    """Create a mock SQLiteDB instance for service testing."""
+    db = MagicMock()
+    return db
+
+
+@pytest.fixture
+def sample_query():
+    return "อริยสัจ"
+
+
+@pytest.fixture
+def sample_query_misspelled():
+    return "อริยะสัจ"

webapp/tipitaka-api/tests/test_pali_utils.py

@@ -0,0 +1,106 @@
+"""
+Unit tests for pali_utils — Thai/Pali text utilities.
+
+Tests cover:
+- Thai digit conversion
+- Pali confusion map variant generation
+- PyThaiNLP autocorrect (with fallback when library is absent)
+- Pali correction dictionary lookup
+- Full correction dict generation
+"""
+
+import pytest
+from app.services.pali_utils import (
+    to_thai_digits,
+    THAI_PALI_CONFUSIONS,
+    _generate_pali_variants,
+    PALI_CORRECTIONS,
+    PALI_CORRECTIONS_FULL,
+    POPULAR_TERMS,
+    HAS_PYTHAINLP,
+    _similar_enough,
+)
+
+
+class TestToThaiDigits:
+    def test_converts_arabic_to_thai(self):
+        assert to_thai_digits("123") == "๑๒๓"
+
+    def test_handles_empty(self):
+        assert to_thai_digits("") == ""
+
+    def test_handles_mixed(self):
+        assert to_thai_digits("มรรค8") == "มรรค๘"
+
+    def test_leaves_thai_digits_unchanged(self):
+        assert to_thai_digits("๑๒๓") == "๑๒๓"
+
+
+class TestPaliConfusions:
+    def test_confusion_map_has_expected_keys(self):
+        assert "ส" in THAI_PALI_CONFUSIONS
+        assert "ณ" in THAI_PALI_CONFUSIONS
+        assert "ภ" in THAI_PALI_CONFUSIONS
+
+    def test_generate_variants(self):
+        variants = _generate_pali_variants("สติ")
+        assert "ศติ" in variants or "ษติ" in variants
+
+    def test_generate_variants_no_confusion(self):
+        """Words without confusion characters should produce no variants."""
+        variants = _generate_pali_variants("ธรรม")
+        assert len(variants) == 0 or all(v != "ธรรม" for v in variants)
+
+
+class TestPaliCorrections:
+    def test_basic_correction(self):
+        assert PALI_CORRECTIONS["อริยะสัจ"] == "อริยสัจ"
+
+    def test_niphan_variants(self):
+        assert PALI_CORRECTIONS["นิพพาณ"] == "นิพพาน"
+        assert PALI_CORRECTIONS["นิพาน"] == "นิพพาน"
+
+    def test_full_dict_includes_hand_curated(self):
+        assert PALI_CORRECTIONS_FULL["อริยะสัจ"] == "อริยสัจ"
+        assert PALI_CORRECTIONS_FULL["นิพพาณ"] == "นิพพาน"
+
+    def test_full_dict_includes_auto_generated(self):
+        """POPULAR_TERMS should have auto-generated variants via confusion map."""
+        if "อริยสัจ" in POPULAR_TERMS:
+            has_variant = any(
+                k != "อริยสัจ" and PALI_CORRECTIONS_FULL.get(k) == "อริยสัจ"
+                for k in PALI_CORRECTIONS_FULL
+            )
+            assert has_variant, "Expected auto-generated variants for อริยสัจ"
+
+    def test_to_thai_digits_in_corrections(self):
+        """Ensure numbered terms have both digit formats."""
+        assert "ขัน5" in PALI_CORRECTIONS
+        assert PALI_CORRECTIONS["ขัน5"] == "ขันธ์ ๕"
+
+
+class TestSimilarEnough:
+    def test_identical(self):
+        assert _similar_enough("สติปัฏฐาน", "สติปัฏฐาน") is True
+
+    def test_similar_passes(self):
+        # Minor variation should pass
+        assert _similar_enough("สติปัฏฐาน", "สติปัฎฐาน") is True
+
+    def test_vastly_different_fails(self):
+        # Very different should fail
+        assert _similar_enough("ปุพเพนิวาสานุสสติญาณ", "ปพเนวสสตญ") is False
+
+    def test_empty(self):
+        assert _similar_enough("", "") is True
+
+
+class TestPopularTerms:
+    def test_popular_terms_is_list_of_strings(self):
+        assert isinstance(POPULAR_TERMS, list)
+        assert len(POPULAR_TERMS) > 0
+        assert all(isinstance(t, str) for t in POPULAR_TERMS)
+
+    def test_includes_core_doctrines(self):
+        core = {"อริยสัจ", "นิพพาน", "ไตรลักษณ์", "สติปัฏฐาน"}
+        assert core.issubset(set(POPULAR_TERMS))

webapp/tipitaka-api/tests/test_rag_service.py

@@ -0,0 +1,59 @@
+"""
+Unit tests for RAGService — embedding cache, Qdrant client.
+
+Uses mocked dependencies to test in isolation.
+"""
+
+import pytest
+from unittest.mock import MagicMock, patch
+
+
+class TestEmbeddingCache:
+    @pytest.fixture
+    def rag_service(self):
+        from app.services.rag_service import RAGService
+        with patch.object(RAGService, '_load_model'), \
+             patch.object(RAGService, '_load_reranker'), \
+             patch.object(RAGService, '_init_qdrant'):
+            service = RAGService()
+            service.client = MagicMock()
+            return service
+
+    def test_cache_initialized(self, rag_service):
+        assert rag_service._embed_cache is not None
+        assert rag_service._embed_cache_max == 256
+
+    def test_cache_stores_and_returns(self, rag_service):
+        """Cache should store and return the same value for identical input."""
+        mock_response = MagicMock()
+        mock_response.json.return_value = {"embeddings": [[0.1, 0.2, 0.3]]}
+        mock_response.raise_for_status.return_value = None
+
+        with patch("app.services.rag_service.httpx.post", return_value=mock_response):
+            emb1 = rag_service._get_embedding("test query")
+
+        # Second call should use cache (no httpx call)
+        emb2 = rag_service._get_embedding("test query")
+        assert emb1 == emb2
+
+    def test_cache_max_respected(self, rag_service):
+        """When cache exceeds max, oldest entries should be evicted."""
+        old_entry = "a" * 10
+        for i in range(rag_service._embed_cache_max + 5):
+            rag_service._embed_cache[f"key_{i}"] = [0.1] * 1024
+            if len(rag_service._embed_cache) > rag_service._embed_cache_max:
+                rag_service._embed_cache.popitem(last=False)
+
+        assert len(rag_service._embed_cache) <= rag_service._embed_cache_max
+
+
+class TestRAGServiceInit:
+    def test_init_graceful_on_qdrant_failure(self):
+        """RAGService constructor should not crash when Qdrant is unavailable."""
+        from app.services.rag_service import RAGService
+        with patch.object(RAGService, '_load_model'), \
+             patch.object(RAGService, '_load_reranker'), \
+             patch.object(RAGService, '_init_qdrant') as mock_init:
+            service = RAGService()
+            mock_init.assert_called_once()
+            assert service.client is None  # _init_qdrant didn't set it

webapp/tipitaka-api/tests/test_search_service.py

@@ -0,0 +1,92 @@
+"""
+Unit tests for SearchService — FTS5, LIKE fallback, autocorrect pipeline.
+
+Uses mocked SQLiteDB to test search logic in isolation.
+"""
+
+import pytest
+from unittest.mock import MagicMock, patch
+from app.services.search_service import SearchService
+from app.services.pali_utils import to_thai_digits
+
+
+class TestCleanText:
+    def setup_method(self):
+        self.service = SearchService(MagicMock())
+
+    def test_removes_line_numbers(self):
+        assert "ข้อความ" in self.service.clean_text("001 ข้อความ")
+
+    def test_removes_footnote_lines(self):
+        assert self.service.clean_text("@เชิงอรรถ") == ""
+
+    def test_removes_separators(self):
+        assert self.service.clean_text("_____") == ""
+
+    def test_preserves_normal_text(self):
+        text = "อริยสัจสี่ ประการ"
+        assert self.service.clean_text(text) == text
+
+    def test_handles_empty(self):
+        assert self.service.clean_text("") == ""
+        assert self.service.clean_text(None) == ""
+
+
+class TestMakeSnippet:
+    def setup_method(self):
+        self.service = SearchService(MagicMock())
+
+    def test_finds_keyword_and_returns_context(self):
+        text = "ดูกรภิกษุทั้งหลาย อริยสัจสี่ ประการเหล่านี้"
+        result = self.service.make_snippet(text, "อริยสัจ")
+        assert "อริยสัจ" in result
+
+    def test_max_length_repsected(self):
+        text = " ".join(["คำ"] * 100)
+        result = self.service.make_snippet(text, "คำ", max_len=100)
+        assert len(result) <= 100 + 20  # allow small expansion
+
+
+class TestHighlight:
+    def setup_method(self):
+        self.service = SearchService(MagicMock())
+
+    def test_wraps_keyword_in_mark(self):
+        result = self.service.highlight("อริยสัจสี่", "อริยสัจ")
+        assert "<mark>อริยสัจ</mark>" in result
+
+    def test_no_keyword_no_change(self):
+        result = self.service.highlight("ข้อความธรรมดา", "อริยสัจ")
+        assert result == "ข้อความธรรมดา"
+
+
+class TestAutocorrectPipeline:
+    def setup_method(self):
+        self.service = SearchService(MagicMock())
+
+    def test_clean_text_no_side_effects(self):
+        """Ensure clean_text doesn't modify input."""
+        text = "001 ข้อความปกติ"
+        result = self.service.clean_text(text)
+        assert result == "ข้อความปกติ"
+
+    def test_search_called_with_corrected_query(self):
+        """Verify search method exists and runs (mocked DB will raise)."""
+        with pytest.raises(Exception):
+            # Will fail on DB mock because connection context manager
+            # but confirms the method signature is correct
+            self.service.search("อริยสัจ", limit=5)
+
+
+class TestLogSearch:
+    def setup_method(self):
+        self.mock_db = MagicMock()
+        self.service = SearchService(self.mock_db)
+
+    def test_empty_query_skipped(self):
+        self.service.log_search("")
+        self.mock_db.get_disk_connection.assert_not_called()
+
+    def test_blank_query_skipped(self):
+        self.service.log_search("   ")
+        self.mock_db.get_disk_connection.assert_not_called()