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

All 1,742 pages with (ย่อ) markers are now pre-indexed in DB.
Remove expensive DeepSeek API call on every (ย่อ) click.
Fix expand_abbreviation to parse plain string arrays from preindex_abbrevs.

scripts/preindex_abbrevs.py

@@ -1,18 +1,22 @@
 """
-Pre-index all (ย่อ) abbreviation markers using NVIDIA NIM (free DeepSeek, 40 RPM).
+Pre-index all (ย่อ) abbreviation markers.
+
+Supports:
+  - NVIDIA NIM (free DeepSeek, 40 RPM)  → --provider nim
+  - DeepSeek API directly               → --provider deepseek
 
 Usage:
-    # Full run (all 1,743 pages)
+    # Full run (default: NVIDIA NIM)
     python scripts/preindex_abbrevs.py
 
-    # Resume from interrupted run
-    python scripts/preindex_abbrevs.py --resume
+    # Resume with DeepSeek API
+    python scripts/preindex_abbrevs.py --resume --provider deepseek
 
     # Test: first 10 pages
     python scripts/preindex_abbrevs.py --limit 10
 
 Scans all pages in tipitaka_mcu.db for (ย่อ) markers.
-For each page, calls NIM DeepSeek to expand ALL abbreviations in one request.
+For each page, calls LLM to expand ALL abbreviations in one request.
 Stores JSON array in reference_markers (type='abbrev', marker_id='_default').
 Supports resume — skips pages already in DB.
 
@@ -43,9 +47,16 @@ ENV_PATH = PROJECT_ROOT.parent / "tipitaka_context" / "nvidia" / ".env"
 # ══════════════════════════════════════════════════════════════
 load_dotenv(ENV_PATH)
 
+# NVIDIA NIM
 NIM_API_KEY = os.getenv("TIPITAKA_API_KEY") or os.getenv("NVIDIA_API_KEY") or ""
 NIM_BASE_URL = os.getenv("TIPITAKA_BASE_URL", "https://integrate.api.nvidia.com/v1")
 NIM_MODEL = os.getenv("TIPITAKA_MODEL", "deepseek-ai/deepseek-v4-flash")
+
+# DeepSeek API
+DEEPSEEK_API_KEY = os.getenv("DEEPSEEK_API_KEY") or os.getenv("DEEPSEEK_API_KEY") or ""
+DEEPSEEK_BASE_URL = os.getenv("DEEPSEEK_BASE_URL", "https://api.deepseek.com")
+DEEPSEEK_MODEL = os.getenv("DEEPSEEK_MODEL", "deepseek-chat")
+
 RATE_INTERVAL = float(os.getenv("TIPITAKA_RATE_INTERVAL", "1.6"))  # 40 RPM
 
 MAX_RETRIES = 4
@@ -95,7 +106,7 @@ def build_prompt(content_text: str) -> str:
 # Parsing
 # ══════════════════════════════════════════════════════════════
 def parse_response(raw: str, expected_count: int) -> list[str]:
-    """Parse JSON array from NIM response into list of expansion strings."""
+    """Parse JSON array from LLM response into list of expansion strings."""
     raw = re.sub(r'<think>.*?</think>', '', raw, flags=re.DOTALL).strip()
     raw = re.sub(r'^```(?:json)?\s*', '', raw)
     raw = re.sub(r'\s*```$', '', raw)
@@ -136,6 +147,7 @@ def parse_response(raw: str, expected_count: int) -> list[str]:
 # ══════════════════════════════════════════════════════════════
 async def expand_page(
     client: AsyncOpenAI,
+    model: str,
     vol: int, page: int, content_text: str,
     conn: sqlite3.Connection,
 ) -> bool:
@@ -149,7 +161,7 @@ async def expand_page(
     for attempt in range(MAX_RETRIES):
         try:
             resp = await client.chat.completions.create(
-                model=NIM_MODEL,
+                model=model,
                 messages=[
                     {"role": "system", "content": SYSTEM_PROMPT},
                     {"role": "user", "content": prompt},
@@ -206,24 +218,45 @@ async def expand_page(
 # Main
 # ══════════════════════════════════════════════════════════════
 async def main():
-    parser = argparse.ArgumentParser(description="Pre-index (ย่อ) abbreviations via NVIDIA NIM")
+    parser = argparse.ArgumentParser(
+        description="Pre-index (ย่อ) abbreviations via LLM"
+    )
     parser.add_argument("--resume", action="store_true", help="Skip already-indexed pages")
     parser.add_argument("--force", action="store_true", help="Re-index even if already in DB")
     parser.add_argument("--limit", type=int, default=0, help="Max pages (0 = all)")
+    parser.add_argument(
+        "--provider", choices=["nim", "deepseek"], default="nim",
+        help="LLM provider: nim (NVIDIA NIM, default) or deepseek (DeepSeek API)"
+    )
     args = parser.parse_args()
 
-    if not NIM_API_KEY:
-        log.error(f"❌ No API key found. Check env file: {ENV_PATH}")
-        log.error("   Set TIPITAKA_API_KEY or NVIDIA_API_KEY")
+    # ── Select provider config ──
+    if args.provider == "nim":
+        api_key = NIM_API_KEY
+        base_url = NIM_BASE_URL
+        model = NIM_MODEL
+        provider_name = "NVIDIA NIM"
+    else:
+        api_key = DEEPSEEK_API_KEY or os.getenv("DEEPSEEK_API_KEY") or ""
+        base_url = DEEPSEEK_BASE_URL
+        model = DEEPSEEK_MODEL
+        provider_name = "DeepSeek API"
+
+    if not api_key:
+        log.error(f"❌ No API key for provider '{args.provider}'. Check env file: {ENV_PATH}")
+        if args.provider == "nim":
+            log.error("   Set TIPITAKA_API_KEY or NVIDIA_API_KEY")
+        else:
+            log.error("   Set DEEPSEEK_API_KEY")
         return
 
     if not DB_PATH.exists():
         log.error(f"❌ DB not found: {DB_PATH}")
         return
 
-    log.info(f"🚀 preindex_abbrevs | model={NIM_MODEL}")
+    log.info(f"🚀 preindex_abbrevs | provider={provider_name} model={model}")
     log.info(f"   DB: {DB_PATH}")
-    log.info(f"   Key: {NIM_API_KEY[:8]}...")
+    log.info(f"   Key: {api_key[:8]}...")
     log.info(f"   Rate: {RATE_INTERVAL}s")
 
     # ── Load pages ──
@@ -264,8 +297,8 @@ async def main():
 
     # ── Setup client ──
     client = AsyncOpenAI(
-        base_url=NIM_BASE_URL,
-        api_key=NIM_API_KEY,
+        base_url=base_url,
+        api_key=api_key,
         max_retries=0,
     )
 
@@ -290,7 +323,7 @@ async def main():
             if wait > 0:
                 await asyncio.sleep(wait)
 
-        ok = await expand_page(client, vol, page, text, conn)
+        ok = await expand_page(client, model, vol, page, text, conn)
 
         if ok:
             success += 1

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

@@ -136,47 +136,19 @@ class LLMService:
                 if content.startswith('['):
                     try:
                         entries = json.loads(content)
-                        parts = [e.get("expansion", "") for e in entries if e.get("expansion")]
-                        return "\n\n".join(parts) if parts else content
-                    except (json.JSONDecodeError, TypeError):
+                        if isinstance(entries, list):
+                            if entries and isinstance(entries[0], str):
+                                # Stored as plain string array (from preindex_abbrevs)
+                                return "\n\n".join(entries)
+                            # Stored as dict array with "expansion" keys
+                            parts = [e.get("expansion", "") for e in entries if isinstance(e.get("expansion", ""), str)]
+                            if parts:
+                                return "\n\n".join(parts)
+                    except (json.JSONDecodeError, TypeError, AttributeError):
                         pass
+                    return content  # raw JSON fallback
                 return content
 
-        # 2. LLM Fallback
-        clean_ctx = self.search_service.clean_text(context) if context else ""
-        prompt = (
-            f"คุณคือผู้เชี่ยวชาญพระไตรปิฎก มจร. "
-            f"ในเนื้อหาต่อไปนี้มีคำว่า '{abbrev_text}' (เช่น (ย่อ), ฯลฯ) "
-            f"กรุณาอธิบายหรือขยายความสิ่งที่ถูกย่อไว้ให้สมบูรณ์ที่สุดตามหลักฐานในพระไตรปิฎก "
-            f"โดยพิจารณาจากบริบทแวดล้อมที่ให้มา:\n\n"
-            f"เนื้อหาบริบท:\n{clean_ctx[:4000]}\n\n"
-            f"กรุณาตอบเป็นข้อความสั้นๆ ที่เป็นเนื้อหาที่ถูกย่อไว้ หรืออธิบายว่าส่วนนี้ย่อมาจากอะไร "
-            f"ถ้าไม่แน่ใจให้บอกว่าเป็นการย่อเพื่อละเนื้อหาที่ซ้ำกัน"
-        )
-        
-        try:
-            response = await self.client.chat.completions.create(
-                model=self.model_map["fast"],
-                messages=[
-                    {"role": "system", "content": "คุณคือผู้เชี่ยวชาญพระไตรปิฎก"},
-                    {"role": "user", "content": prompt}
-                ],
-                max_tokens=500
-            )
-            result = response.choices[0].message.content.strip()
-            
-            # 3. Cache result in DB for future lookups
-            try:
-                with self.db.get_connection() as conn:
-                    cursor = conn.cursor()
-                    cursor.execute("""
-                        INSERT OR REPLACE INTO reference_markers
-                            (volume_num, page_num, marker_id, type, content)
-                        VALUES (?, ?, ?, 'abbrev', ?)
-                    """, (vol, page, lookup_id, result))
-            except Exception:
-                pass  # Cache failure is non-critical
-            
-            return result
-        except Exception as e:
-            return f"ไม่สามารถขยายความได้ในขณะนี้: {str(e)}"
+        # 2. LLM Fallback — DISABLED (all abbrevs pre-indexed in DB)
+        # If somehow missing, return a default message instead of calling AI
+        return f"(ย่อ) — ยังไม่มีข้อมูลขยายความในฐานข้อมูลเล่ม {vol} หน้า {page}"