fix(litellm): fall back to last user message when hindsight_query not provided (#641)

* fix(litellm): fall back to last user message when hindsight_query not provided

inject_memories=True no longer requires an explicit hindsight_query. The
injection path now falls back to extracting the last user message, matching
the documented Quick Start behavior that was broken since #167 (v0.4.18).

* test(litellm): add regression tests for inject_memories without hindsight_query

hindsight-integrations/litellm/hindsight_litellm/__init__.py

@@ -262,15 +262,28 @@ def _inject_memories(
     if not messages:
         return messages
 
-    # hindsight_query is required when inject_memories=True
-    if not custom_query:
-        raise ValueError(
-            "hindsight_query is required when inject_memories=True. "
-            "Pass hindsight_query='your query' to specify what to search for in memory. "
-            "Example: hindsight_query=recipient_name or hindsight_query='What do I know about Alice?'"
-        )
-
-    user_query = custom_query
+    # Use custom_query if provided, otherwise fall back to the last user message
+    if custom_query:
+        user_query = custom_query
+    else:
+        user_query = None
+        for msg in reversed(messages):
+            if msg.get("role") == "user":
+                content = msg.get("content")
+                if isinstance(content, str):
+                    user_query = content
+                    break
+                elif isinstance(content, list):
+                    text_parts = [
+                        item.get("text", "")
+                        for item in content
+                        if isinstance(item, dict) and item.get("type") == "text"
+                    ]
+                    if text_parts:
+                        user_query = " ".join(text_parts)
+                        break
+        if not user_query:
+            return messages
 
     # Use bank_id from defaults
     bank_id = defaults.bank_id

hindsight-integrations/litellm/hindsight_litellm/callbacks.py

@@ -162,9 +162,7 @@ class HindsightCallback(CustomLogger):
                 logger.error(f"HTTP POST failed: {e}")
             raise HindsightError(f"Hindsight API request failed: {e}") from e
 
-    def _http_get(
-        self, url: str, config: HindsightConfig
-    ) -> Optional[dict]:
+    def _http_get(self, url: str, config: HindsightConfig) -> Optional[dict]:
         """Make a synchronous HTTP GET request.
 
         Returns:
@@ -736,16 +734,11 @@ class HindsightCallback(CustomLogger):
         if self._should_skip_model(model, config):
             return
 
-        # hindsight_query is required when inject_memories=True
+        # Use hindsight_query if provided, otherwise fall back to the last user message
         custom_query = kwargs.get("hindsight_query")
-        if not custom_query:
-            raise ValueError(
-                "hindsight_query is required when inject_memories=True. "
-                "Pass hindsight_query='your query' to specify what to search for in memory. "
-                "Example: hindsight_query=recipient_name or hindsight_query='What do I know about Alice?'"
-            )
-
-        user_query = custom_query
+        user_query = custom_query or self._extract_user_query(messages)
+        if not user_query:
+            return
 
         # Use reflect or recall based on settings
         if settings.use_reflect:
@@ -802,16 +795,11 @@ class HindsightCallback(CustomLogger):
         if self._should_skip_model(model, config):
             return
 
-        # hindsight_query is required when inject_memories=True
+        # Use hindsight_query if provided, otherwise fall back to the last user message
         custom_query = kwargs.get("hindsight_query")
-        if not custom_query:
-            raise ValueError(
-                "hindsight_query is required when inject_memories=True. "
-                "Pass hindsight_query='your query' to specify what to search for in memory. "
-                "Example: hindsight_query=recipient_name or hindsight_query='What do I know about Alice?'"
-            )
-
-        user_query = custom_query
+        user_query = custom_query or self._extract_user_query(messages)
+        if not user_query:
+            return
 
         # Use reflect or recall based on settings
         if settings.use_reflect:

hindsight-integrations/litellm/tests/test_integration.py

@@ -356,6 +356,74 @@ class TestCallback:
         assert "Relevant Memories" in result[0]["content"]
         assert "What's my name?" in result[0]["content"]
 
+    def test_inject_memories_uses_last_user_message_when_no_hindsight_query(self):
+        """Regression test: inject_memories=True should not require hindsight_query.
+
+        The documented Quick Start example does not pass hindsight_query; the
+        injection path must fall back to the last user message automatically.
+        See: feat(litellm) #167 regression.
+        """
+        from unittest.mock import MagicMock, patch
+
+        callback = HindsightCallback()
+
+        configure(
+            hindsight_api_url="http://localhost:8888",
+            inject_memories=True,
+        )
+        set_defaults(bank_id="test-agent")
+
+        messages = [{"role": "user", "content": "What did we discuss about AI?"}]
+        kwargs = {}  # No hindsight_query provided — this is the regression scenario
+
+        mock_memory = MagicMock()
+        mock_memory.text = "AI is cool"
+        mock_memory.type = "world"
+        mock_memory.weight = 0.9
+
+        with patch.object(callback, "_recall_memories_sync", return_value=[mock_memory]) as mock_recall:
+            callback.log_pre_api_call(
+                model="gpt-4o-mini",
+                messages=messages,
+                kwargs=kwargs,
+            )
+            # Should have called recall with the last user message as query
+            mock_recall.assert_called_once()
+            query_used = mock_recall.call_args[0][0]
+            assert query_used == "What did we discuss about AI?"
+
+        # Memories should have been injected into messages
+        assert any("AI is cool" in str(m.get("content", "")) for m in messages)
+
+    def test_inject_memories_hindsight_query_takes_precedence(self):
+        """When hindsight_query is provided it should be used over the last user message."""
+        from unittest.mock import MagicMock, patch
+
+        callback = HindsightCallback()
+
+        configure(
+            hindsight_api_url="http://localhost:8888",
+            inject_memories=True,
+        )
+        set_defaults(bank_id="test-agent")
+
+        messages = [{"role": "user", "content": "Hello"}]
+        kwargs = {"hindsight_query": "What do I know about Alice?"}
+
+        mock_memory = MagicMock()
+        mock_memory.text = "Alice likes cats"
+        mock_memory.type = "world"
+        mock_memory.weight = 0.9
+
+        with patch.object(callback, "_recall_memories_sync", return_value=[mock_memory]) as mock_recall:
+            callback.log_pre_api_call(
+                model="gpt-4o-mini",
+                messages=messages,
+                kwargs=kwargs,
+            )
+            query_used = mock_recall.call_args[0][0]
+            assert query_used == "What do I know about Alice?"
+
     def test_should_skip_model_exact_match(self):
         """Test model exclusion with exact match."""
         callback = HindsightCallback()