ollama側のレスポンスが空の場合に1回リトライする処理に変更

トークン数の累積計算処理削除

6 files changed, 74 insertions, 15 deletions

diff --git a/src/ollama_web/i18n/messages/en.json b/src/ollama_web/i18n/messages/en.json
index 9e2894a..42bce33 100644
--- a/src/ollama_web/i18n/messages/en.json
+++ b/src/ollama_web/i18n/messages/en.json
@@ -78,7 +78,9 @@
     "tool_limit_reached": "Tool call limit reached. Summarizing findings…",
     "thinking": "Thinking process",
     "executing_tool": "Executing {label}…",
-    "running": "Running…"
+    "running": "Running…",
+    "empty_response": "The model returned an empty response. Please send the message again.",
+    "empty_response_retry": "Empty response received; retrying…"
   },
   "errors": {
     "generic": "An error occurred",
diff --git a/src/ollama_web/i18n/messages/ja.json b/src/ollama_web/i18n/messages/ja.json
index aa9f01e..a476ca2 100644
--- a/src/ollama_web/i18n/messages/ja.json
+++ b/src/ollama_web/i18n/messages/ja.json
@@ -78,7 +78,9 @@
     "tool_limit_reached": "ツール呼び出し上限に達しました。これまでの調査結果を整理中…",
     "thinking": "思考プロセス",
     "executing_tool": "{label} を実行中…",
-    "running": "実行中…"
+    "running": "実行中…",
+    "empty_response": "モデルから空の応答が返されました。もう一度送信してください。",
+    "empty_response_retry": "応答が空でした。リトライ中…"
   },
   "errors": {
     "generic": "エラーが発生しました",
diff --git a/src/ollama_web/llm.py b/src/ollama_web/llm.py
index 48c177a..76a61a5 100644
--- a/src/ollama_web/llm.py
+++ b/src/ollama_web/llm.py
@@ -393,6 +393,9 @@ def stream_chat_with_tools(
 
     cumulative_tool_chars = 0
 
+    MAX_EMPTY_RETRIES = 2
+    empty_retries = 0
+
     rounds = 0
     while rounds < MAX_TOOL_ROUNDS:
         rounds += 1
@@ -406,6 +409,7 @@ def stream_chat_with_tools(
             model,
             cumulative_tool_chars,
         )
+
         t0 = time.time()
         try:
             stream = cast(Any, raw_client).chat(
@@ -449,6 +453,35 @@ def stream_chat_with_tools(
             time.time() - t0,
         )
 
+        # Detect empty responses. Some models / proxies occasionally return a
+        # stream with almost no meaningful content. Log it and retry a few
+        # times before giving up.
+        is_empty = (
+            not full_message.get("content")
+            and not full_message.get("thinking")
+            and not tool_calls
+        )
+        if is_empty and empty_retries < MAX_EMPTY_RETRIES:
+            empty_retries += 1
+            logger.warning(
+                "ollama returned an empty stream (chunks=%d); retry %d/%d",
+                chunk_count,
+                empty_retries,
+                MAX_EMPTY_RETRIES,
+            )
+            yield {
+                "type": "status",
+                "message": t("status.empty_response_retry", lang=language),
+            }
+            # Retry the same round without consuming the tool round budget.
+            rounds -= 1
+            # Small pause to avoid hammering the upstream.
+            time.sleep(0.5)
+            continue
+        # Reset retry budget on any non-empty round so isolated hiccups are fine.
+        if not is_empty:
+            empty_retries = 0
+
         # Reconstruct the assistant message including any tool_calls so the
         # next round can see them.
         if tool_calls:
@@ -458,6 +491,12 @@ def stream_chat_with_tools(
             messages.append(msg)
 
         if not tool_calls:
+            if is_empty:
+                # Ran out of tool rounds while still empty; surface a warning.
+                yield {
+                    "type": "warning",
+                    "message": t("status.empty_response", lang=language),
+                }
             done_event = {"type": "done"}
             tokens = _token_counts(last_chunk)
             if tokens:
diff --git a/src/ollama_web/sessions.py b/src/ollama_web/sessions.py
index 593f2d7..388066b 100644
--- a/src/ollama_web/sessions.py
+++ b/src/ollama_web/sessions.py
@@ -243,19 +243,23 @@ class SessionStore:
         return session
 
     def _update_session_token_count(self, session: dict[str, Any]) -> None:
-        """Recalculate cumulative token_count for the session."""
-        total_prompt = 0
-        total_eval = 0
-        for m in session.get("messages", []):
-            tc = m.get("token_count") or {}
-            total_prompt += int(tc.get("prompt_eval_count", 0) or 0)
-            total_eval += int(tc.get("eval_count", 0) or 0)
-        if total_prompt or total_eval:
-            session["token_count"] = {
-                "prompt_eval_count": total_prompt,
-                "eval_count": total_eval,
-                "total_count": total_prompt + total_eval,
-            }
+        """Update the session token_count from the most recent assistant message.
+
+        ``prompt_eval_count`` is the input token count for the *entire* conversation
+        as seen by ollama in that response, not a per-message value.  Summing it
+        across messages would over-count massively.  We therefore keep the latest
+        assistant message's token_count as the authoritative session value.
+        """
+        messages = session.get("messages", [])
+        latest_assistant_tc = None
+        for m in reversed(messages):
+            if m.get("role") == "assistant":
+                tc = m.get("token_count")
+                if tc:
+                    latest_assistant_tc = tc
+                    break
+        if latest_assistant_tc is not None:
+            session["token_count"] = dict(latest_assistant_tc)
         else:
             session.pop("token_count", None)
 
diff --git a/src/ollama_web/static/app.js b/src/ollama_web/static/app.js
index 378e39f..54fb2e0 100644
--- a/src/ollama_web/static/app.js
+++ b/src/ollama_web/static/app.js
@@ -714,6 +714,17 @@ async function send() {
             waiting.textContent = ev.message || t("chat.processing");
             assistantBody.appendChild(waiting);
           }
+        } else if (ev.type === "warning") {
+          gotError = true;
+          if (!placeholderCleared) {
+            assistantBody.innerHTML = "";
+            placeholderCleared = true;
+          }
+          const span = document.createElement("span");
+          span.className = "warning-msg";
+          span.style.color = "#f0883e";
+          span.textContent = ev.message || t("status.empty_response");
+          assistantBody.appendChild(span);
         } else if (ev.type === "thinking") {
           if (!thinkingBody) thinkingBody = addThinkingBubble(assistantExtras);
           thinkingBody.textContent += ev.content;
diff --git a/tests/test_clear_messages.py b/tests/test_clear_messages.py
index c3f3595..a0653ca 100644
--- a/tests/test_clear_messages.py
+++ b/tests/test_clear_messages.py
@@ -82,6 +82,7 @@ def test_add_message_stores_token_count_and_clears_it():
     assert reloaded["token_count"] == {
         "prompt_eval_count": 120,
         "eval_count": 80,
+        "num_ctx": 8192,
         "total_count": 200,
     }