Add web search knowledge base, message tagging, and word-memory tip topics

Introduces websearch.py (Tavily) + knowledge.py to ground chat responses
in search results, tagging.py to tag user and assistant messages via LLM
and pick related topics for the Discord useless-tips job, and
word_memory.py to track recent conversation nouns (surfaced in the new
WordPool.vue sidebar) used to pick tip topics close to a related tag.

backend/app/core/llm.py

@@ -7,6 +7,7 @@ from contextlib import asynccontextmanager
 
 import litellm
 from app.core.config import DEFAULT_MODEL, DEBUG
+from app.core.websearch import WEB_SEARCH_TOOL, search_web
 
 litellm.drop_params = True
 if DEBUG:
@@ -58,6 +59,7 @@ async def stream_chat(
     messages: list[dict],
     model: str | None = None,
     session_id: str | None = None,
+    tools: list[dict] | None = None,
 ) -> AsyncIterator[str]:
     used_model = model or DEFAULT_MODEL
     async with _log_context(used_model, messages, session_id) as log:
@@ -65,6 +67,7 @@ async def stream_chat(
             model=used_model,
             messages=messages,
             stream=True,
+            **({"tools": tools} if tools else {}),
         )
         chunks: list[str] = []
         async for chunk in response:
@@ -92,3 +95,70 @@ async def complete(
         log["response_text"] = result
     _save_log(log)
     return result
+
+
+async def resolve_web_search(
+    messages: list[dict],
+    model: str | None = None,
+    session_id: str | None = None,
+    max_rounds: int = 2,
+) -> tuple[list[dict], list[dict]]:
+    """必要に応じてLLMにweb_searchツールを呼ばせ、検索結果を会話に組み込んだ上でメッセージ列を返す。
+
+    モデルがtool callingに対応していない場合は何もせず元のmessagesをそのまま返す。
+    戻り値: (拡張後のmessages, 実際に検索したクエリと結果のリスト)
+    """
+    used_model = model or DEFAULT_MODEL
+    working_messages = list(messages)
+    searches: list[dict] = []
+
+    for _ in range(max_rounds):
+        try:
+            response = await litellm.acompletion(
+                model=used_model,
+                messages=working_messages,
+                tools=[WEB_SEARCH_TOOL],
+                tool_choice="auto",
+            )
+        except Exception:
+            logger.info("Model %s does not support tool calling; skipping web search", used_model)
+            return messages, searches
+
+        choice_message = response.choices[0].message
+        tool_calls = getattr(choice_message, "tool_calls", None)
+        if not tool_calls:
+            break
+
+        working_messages.append({
+            "role": "assistant",
+            "content": choice_message.content or "",
+            "tool_calls": [
+                {
+                    "id": tc.id,
+                    "type": "function",
+                    "function": {"name": tc.function.name, "arguments": tc.function.arguments},
+                }
+                for tc in tool_calls
+            ],
+        })
+
+        for tc in tool_calls:
+            try:
+                args = json.loads(tc.function.arguments or "{}")
+            except json.JSONDecodeError:
+                args = {}
+            query = args.get("query", "")
+            results = await search_web(query)
+            if results:
+                tool_text = "\n\n".join(f"[{r['title']}]({r['url']})\n{r['content']}" for r in results)
+                searches.append({"query": query, "results": results})
+            else:
+                tool_text = "検索結果が見つかりませんでした。"
+
+            working_messages.append({
+                "role": "tool",
+                "tool_call_id": tc.id,
+                "content": tool_text,
+            })
+
+    return working_messages, searches