vision-agent 0.2.166__tar.gz → 0.2.167__tar.gz

{vision_agent-0.2.166 → vision_agent-0.2.167}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: vision-agent
-Version: 0.2.166
+Version: 0.2.167
 Summary: Toolset for Vision Agent
 Author: Landing AI
 Author-email: dev@landing.ai

{vision_agent-0.2.166 → vision_agent-0.2.167}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "vision-agent"
-version = "0.2.166"
+version = "0.2.167"
 description = "Toolset for Vision Agent"
 authors = ["Landing AI <dev@landing.ai>"]
 readme = "README.md"

{vision_agent-0.2.166 → vision_agent-0.2.167}/vision_agent/agent/vision_agent.py

@@ -85,6 +85,15 @@ def format_agent_message(agent_message: str) -> str:
     return output
 
 
+def _clean_response(response: str) -> str:
+    # Sometimes the LLM will hallucinate responses to an <execute_python> tag as if it
+    # had already executed the code. This function removes the hallucinated response.
+    if "<execute_python>" in response:
+        end_execute_python = response.find("</execute_python>")
+        response = response[: end_execute_python + len("</execute_python>")]
+    return response
+
+
 def run_conversation(orch: LMM, chat: List[Message]) -> Dict[str, Any]:
     chat = copy.deepcopy(chat)
 
@@ -114,6 +123,10 @@ def run_conversation(orch: LMM, chat: List[Message]) -> Dict[str, Any]:
         message["media"] = chat[-1]["media"]
     conv_resp = cast(str, orch([message], stream=False))
 
+    # clean the response first, if we are executing code, do not resond or end
+    # conversation before the code has been executed.
+    conv_resp = _clean_response(conv_resp)
+
     let_user_respond_str = extract_tag(conv_resp, "let_user_respond")
     let_user_respond = (
         "true" in let_user_respond_str.lower() if let_user_respond_str else False
@@ -458,7 +471,7 @@ class VisionAgent(Agent):
                     self.streaming_message(
                         {
                             "role": "assistant",
-                            "content": json.dumps(response),
+                            "content": json.dumps(add_step_descriptions(response)),
                             "finished": finished and code_action is None,
                         }
                     )

{vision_agent-0.2.166 → vision_agent-0.2.167}/vision_agent/tools/meta_tools.py

@@ -676,12 +676,13 @@ def use_extra_vision_agent_args(
     for node in red:
         # seems to always be atomtrailers not call type
         if node.type == "atomtrailers":
+            if node.name.value == "generate_vision_code":
+                node.value[1].value.append(f"test_multi_plan={test_multi_plan}")
+
             if (
                 node.name.value == "generate_vision_code"
                 or node.name.value == "edit_vision_code"
             ):
-                node.value[1].value.append(f"test_multi_plan={test_multi_plan}")
-
                 if custom_tool_names is not None:
                     node.value[1].value.append(f"custom_tool_names={custom_tool_names}")
     cleaned_code = red.dumps().strip()