Update output_parser.py (#10430)

- Description: Updated output parser for mrkl to remove any
hallucination actions after the final answer; this was encountered when
using Anthropic claude v2 for planning; reopening PR with updated unit
tests
- Issue: #10278 
- Dependencies: N/A
- Twitter handle: @johnreynolds

libs/langchain/langchain/agents/mrkl/output_parser.py

@@ -29,11 +29,20 @@ class MRKLOutputParser(AgentOutputParser):
             r"Action\s*\d*\s*:[\s]*(.*?)[\s]*Action\s*\d*\s*Input\s*\d*\s*:[\s]*(.*)"
         )
         action_match = re.search(regex, text, re.DOTALL)
-        if action_match:
-            if includes_answer:
+        if action_match and includes_answer:
+            if text.find(FINAL_ANSWER_ACTION) < text.find(action_match.group(0)):
+                # if final answer is before the hallucination, return final answer
+                start_index = text.find(FINAL_ANSWER_ACTION) + len(FINAL_ANSWER_ACTION)
+                end_index = text.find("\n\n", start_index)
+                return AgentFinish(
+                    {"output": text[start_index:end_index].strip()}, text[:end_index]
+                )
+            else:
                 raise OutputParserException(
                     f"{FINAL_ANSWER_AND_PARSABLE_ACTION_ERROR_MESSAGE}: {text}"
                 )
+
+        if action_match:
             action = action_match.group(1).strip()
             action_input = action_match.group(2)
             tool_input = action_input.strip(" ")

libs/langchain/tests/unit_tests/agents/test_mrkl_output_parser.py

@@ -52,12 +52,26 @@ def test_missing_action_input() -> None:
     )
 
 
-def test_final_answer_and_parsable_action() -> None:
+def test_final_answer_before_parsable_action() -> None:
     llm_output = """Final Answer: The best pizza to eat is margaritta
-        I can use the `foo` tool to achieve the goal.
+
         Action: foo
         Action Input: bar
         """
+    agent_finish: AgentFinish = mrkl_output_parser.parse(llm_output)  # type: ignore
+    assert (
+        agent_finish.return_values.get("output")
+        == "The best pizza to eat is margaritta"
+    )
+
+
+def test_final_answer_after_parsable_action() -> None:
+    llm_output = """
+        Observation: I can use the `foo` tool to achieve the goal.
+        Action: foo
+        Action Input: bar
+        Final Answer: The best pizza to eat is margaritta 
+        """
     with pytest.raises(OutputParserException) as exception_info:
         mrkl_output_parser.parse(llm_output)
     assert (