Update src/api/agent.py

src/api/agent.py

@@ -1750,6 +1750,54 @@ Return ONLY valid JSON:
 {{"corrected":"...","natural":"...","boolean":"(\"A\" OR \"B\") AND (\"C\" OR \"D\")","year_from":"","year_to":"","peer_reviewed":false,"open_access":false}}"""
 
 
+async def _validate_topic(topic: str, model: str) -> Tuple[bool, str]:
+    """
+    Stage 2 topic validator — runs after the regex ambiguity gate.
+    Catches gibberish/unrecognisable words that passed the gate.
+
+    Returns:
+        (True, corrected_topic)  — topic is recognisable; corrected if typo
+        (False, "")              — topic is gibberish/unrecognisable
+
+    Uses a tiny ~30-token LLM call. Fails open (returns True) on any exception
+    so a validator failure never blocks a legitimate search.
+    """
+    settings = get_settings()
+    if not settings.openai_api_key and not settings.anthropic_api_key:
+        return True, topic  # No LLM available — fail open
+    try:
+        llm = _get_llm(model, temperature=0, max_tokens=40)
+        response = await llm.ainvoke([
+            {"role": "system", "content": (
+                "You are a topic validator for an academic library search system. "
+                "Given a word or phrase, decide if it is a recognisable research topic, "
+                "subject area, acronym, proper noun, or concept — even if misspelled.\n\n"
+                "Rules:\n"
+                "- If recognisable or a fixable typo: return JSON {\"valid\": true, \"corrected\": \"<corrected spelling>\"}\n"
+                "- If gibberish, random characters, or completely unrecognisable: return JSON {\"valid\": false, \"corrected\": \"\"}\n"
+                "- Acronyms like NLP, ML, AI, CRISPR, IoT are always valid.\n"
+                "- Proper nouns (country names, people, organisations) are always valid.\n"
+                "- Misspellings like 'machne lernig' → corrected: 'machine learning' → valid: true\n"
+                "- Random strings like 'sdfmdnoc', 'xyzabc123', 'qwerty' → valid: false\n"
+                "Return ONLY valid JSON. No explanation."
+            )},
+            {"role": "user", "content": f"Topic: {topic}"},
+        ])
+        raw = response.content.strip()
+        if raw.startswith("```"):
+            raw = raw.split("\n", 1)[1].rsplit("```", 1)[0].strip()
+        s, e = raw.find("{"), raw.rfind("}")
+        if s != -1 and e > s:
+            result = json.loads(raw[s:e + 1])
+            is_valid = bool(result.get("valid", True))
+            corrected = str(result.get("corrected") or topic).strip()
+            return is_valid, corrected
+        return True, topic  # Malformed response — fail open
+    except Exception as exc:
+        logger.warning(f"_validate_topic failed: {exc} — failing open")
+        return True, topic  # Always fail open so validator never blocks search
+
+
 async def _build_search_plan(question: str, model: str) -> dict:
     """
     LLM-driven search plan — ported from app81.py and adapted for LibBee.
@@ -1796,6 +1844,37 @@ async def _build_search_plan(question: str, model: str) -> dict:
         }
     # ── End ambiguity gate ────────────────────────────────────────────────────
 
+    # ── Stage 2: LLM topic validator ─────────────────────────────────────────
+    # Catches gibberish/unrecognisable words that passed the regex gate.
+    # Typos are silently corrected. Gibberish triggers clarification.
+    # Uses a tiny ~30-token LLM call — much cheaper than running full boolean
+    # builder on garbage input.
+    if settings.openai_api_key or settings.anthropic_api_key:
+        _valid, _corrected_topic = await _validate_topic(_stripped, model)
+        if not _valid:
+            return {
+                "corrected":            raw_query,
+                "natural":              light_query,
+                "boolean":              "",
+                "database_query":       "",
+                "year_from":            "",
+                "year_to":              "",
+                "peer_reviewed":        False,
+                "open_access":          False,
+                "clarification_needed": True,
+                "clarification_message": (
+                    f"I couldn't recognise <strong>{_escape(_stripped)}</strong> as a research topic. "
+                    f"Could you check the spelling, or describe what you're looking for in more detail? "
+                    f"For example: <em>\"find research on machine learning\"</em> or "
+                    f"<em>\"articles on renewable energy in UAE\"</em>."
+                ),
+            }
+        # Silently use corrected spelling for the rest of the pipeline
+        if _corrected_topic and _corrected_topic.lower() != _stripped.lower():
+            logger.info(f"_validate_topic corrected: {_stripped!r} → {_corrected_topic!r}")
+            light_query = _corrected_topic
+    # ── End stage 2 ──────���────────────────────────────────────────────────────
+
     if settings.openai_api_key or settings.anthropic_api_key:
         try:
             llm = _get_llm(model, temperature=0, max_tokens=300)