langroid 0.1.241__tar.gz → 0.1.244__tar.gz

{langroid-0.1.241 → langroid-0.1.244}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: langroid
-Version: 0.1.241
+Version: 0.1.244
 Summary: Harness LLMs with Multi-Agent Programming
 License: MIT
 Author: Prasad Chalasani
@@ -30,7 +30,7 @@ Requires-Dist: chainlit (>=1.0.400,<2.0.0) ; extra == "chainlit"
 Requires-Dist: chromadb (>=0.4.21,<=0.4.23) ; extra == "chromadb"
 Requires-Dist: colorlog (>=6.7.0,<7.0.0)
 Requires-Dist: docstring-parser (>=0.15,<0.16)
-Requires-Dist: duckduckgo-search (>=5.3.0,<6.0.0)
+Requires-Dist: duckduckgo-search (>=6.0.0,<7.0.0)
 Requires-Dist: faker (>=18.9.0,<19.0.0)
 Requires-Dist: fakeredis (>=2.12.1,<3.0.0)
 Requires-Dist: fire (>=0.5.0,<0.6.0)

{langroid-0.1.241 → langroid-0.1.244}/langroid/agent/base.py

@@ -812,12 +812,14 @@ class Agent(ABC):
             llm_model = (
                 "no-LLM" if self.config.llm is None else self.llm.config.chat_model
             )
-
+            # tot cost across all LLMs, agents
+            all_cost = format(self.llm.tot_tokens_cost()[1], ".4f")
             return (
                 f"[bold]Stats:[/bold] [magenta]N_MSG={chat_length}, "
                 f"TOKENS: in={in_tokens}, out={out_tokens}, "
                 f"max={max_out}, ctx={context_length}, "
-                f"COST: now=${llm_response_cost}, cumul=${cumul_cost} "
+                f"COST: now=${llm_response_cost}, cumul=${cumul_cost}, "
+                f"tot=${all_cost} "
                 f"[bold]({llm_model})[/bold][/magenta]"
             )
         return ""

{langroid-0.1.241 → langroid-0.1.244}/langroid/agent/special/lance_rag/critic_agent.py

@@ -52,7 +52,7 @@ class QueryPlanCriticConfig(LanceQueryPlanAgentConfig):
       SEMANTIC/LEXICAL/FUZZY FILTER since the Assistant is able to use it to match 
       the CONTENT of the docs in various ways (semantic, lexical, fuzzy, etc.).
          
-    - DATAFRAME CALCULATION, and 
+    - DATAFRAME CALCULATION, which must be a SINGLE LINE calculation, and 
     - ANSWER recieved from an assistant that used this QUERY PLAN.
 
     In addition to the above SCHEMA fields there is a `content` field which:
@@ -64,8 +64,10 @@ class QueryPlanCriticConfig(LanceQueryPlanAgentConfig):
     ONLY using the `query_plan_feedback` tool, and DO NOT SAY ANYTHING ELSE.
     
     Here is how you must examine the QUERY PLAN + ANSWER:
+    - ALL filtering conditions in the original query must be EXPLICITLY 
+      mentioned in the FILTER, and the QUERY field should not be used for filtering.
     - If the ANSWER is in the expected form, then the QUERY PLAN is likely VALID,
-      and your feedback should be EMPTY.
+      and your feedback should say VALID, with empty `suggested_fix`.
     - If the ANSWER is {NO_ANSWER} or of the wrong form, 
       then try to DIAGNOSE the problem IN THE FOLLOWING ORDER:
       - DATAFRAME CALCULATION -- is it doing the right thing?
@@ -83,11 +85,13 @@ class QueryPlanCriticConfig(LanceQueryPlanAgentConfig):
         REMEMBER: A filter should ONLY be used if EXPLICITLY REQUIRED BY THE QUERY.
      
     
-    ALWAYS use `query_plan_feedback` tool/fn to present your feedback!
-    and DO NOT SAY ANYTHING ELSE OUTSIDE THE TOOL/FN.
-    IF NO REVISION NEEDED, simply give EMPTY FEEBACK, SAY NOTHING ELSE
-    and DO NOT EXPLAIN YOURSELF.
-        
+    ALWAYS use `query_plan_feedback` tool/fn to present your feedback
+    in the `feedback` field, and if any fix is suggested,
+    present it in the `suggested_fix` field.
+    DO NOT SAY ANYTHING ELSE OUTSIDE THE TOOL/FN.
+    IF NO REVISION NEEDED, simply leave the `suggested_fix` field EMPTY,
+    and SAY NOTHING ELSE
+    and DO NOT EXPLAIN YOURSELF.        
     """
 
 
@@ -95,7 +99,7 @@ def plain_text_query_plan(msg: QueryPlanAnswerTool) -> str:
     plan = f"""
     OriginalQuery: {msg.plan.original_query}
     Filter: {msg.plan.filter}
-    Query: {msg.plan.query}
+    Rephrased Query: {msg.plan.query}
     DataframeCalc: {msg.plan.dataframe_calc}
     Answer: {msg.answer}
     """
@@ -126,11 +130,12 @@ class QueryPlanCritic(ChatAgent):
     def handle_message_fallback(
         self, msg: str | ChatDocument
     ) -> str | ChatDocument | None:
-        """Create QueryPlanFeedbackTool since LLM forgot"""
+        """Remind the LLM to use QueryPlanFeedbackTool since it forgot"""
         if isinstance(msg, ChatDocument) and msg.metadata.sender == Entity.LLM:
-            # our LLM forgot to use the QueryPlanFeedbackTool
-            feedback = QueryPlanFeedbackTool(feedback=msg.content)
-            msg.tool_messages = [feedback]
-            msg.content = DONE
-            return msg
+            return """
+            You forgot to use the `query_plan_feedback` tool/function.
+            Re-try your response using the `query_plan_feedback` tool/function,
+            remember to provide feedback in the `feedback` field,
+            and if any fix is suggested, provide it in the `suggested_fix` field.
+            """
         return None

{langroid-0.1.241 → langroid-0.1.244}/langroid/agent/special/lance_rag/query_planner_agent.py

@@ -16,6 +16,7 @@ This agent has access to two tools:
 
 import logging
 
+import langroid as lr
 from langroid.agent.chat_agent import ChatAgent, ChatAgentConfig
 from langroid.agent.chat_document import ChatDocument
 from langroid.agent.special.lance_tools import (
@@ -35,6 +36,7 @@ class LanceQueryPlanAgentConfig(ChatAgentConfig):
     doc_agent_name: str = "LanceRAG"
     doc_schema: str = ""
     use_tools = False
+    max_retries: int = 5  # max number of retries for query plan
     use_functions_api = True
 
     system_message = f"""
@@ -58,10 +60,13 @@ class LanceQueryPlanAgentConfig(ChatAgentConfig):
         so the REPHRASED QUERY should NOT mention ANY FILTER fields.
         The answer will answer based on documents whose CONTENTS match the QUERY, 
         possibly REPHRASED. 
-    - a Pandas-dataframe calculation/aggregation string that can be used to calculate
-        the answer to the original query, e.g. "df["rating"].mean()",
-        or "df.groupby("director").mean()["rating"]", etc, or empty string if no calc
-        is needed. The dataframe calc CAN refer to the `content` field. 
+    - an OPTIONAL SINGLE-LINE Pandas-dataframe calculation/aggregation string 
+        that can be used to calculate the answer to the original query, 
+        e.g. "df["rating"].mean()",
+        or "df.groupby("director").mean()["rating"]", 
+        or EMPTY string if no calc is needed. 
+        The dataframe calc CAN refer to the `content` field.
+        If a DataFrame calculation is NOT needed, leave this field EMPTY.
     
     
     EXAMPLE:
@@ -116,6 +121,8 @@ class LanceQueryPlanAgent(ChatAgent):
         super().__init__(config)
         self.config: LanceQueryPlanAgentConfig = config
         self.curr_query_plan: QueryPlan | None = None
+        # how many times re-trying query plan in response to feedback:
+        self.n_retries: int = 0
         self.result: str = ""  # answer received from LanceRAG
         # This agent should generate the QueryPlanTool
         # as well as handle it for validation
@@ -125,6 +132,8 @@ class LanceQueryPlanAgent(ChatAgent):
     def query_plan(self, msg: QueryPlanTool) -> str:
         """Valid, forward to RAG Agent"""
         # save, to be used to assemble QueryPlanResultTool
+        if len(msg.plan.dataframe_calc.split("\n")) > 1:
+            return "DATAFRAME CALCULATION must be a SINGLE LINE; Retry the `query_plan`"
         self.curr_query_plan = msg.plan
         return PASS_TO + self.config.doc_agent_name
 
@@ -132,7 +141,8 @@ class LanceQueryPlanAgent(ChatAgent):
         """Process Critic feedback on QueryPlan + Answer from RAG Agent"""
         # We should have saved answer in self.result by this time,
         # since this Agent seeks feedback only after receiving RAG answer.
-        if msg.feedback == "":
+        if msg.suggested_fix == "":
+            self.n_retries = 0
             # This means the Query Plan or Result is good, as judged by Critic
             if self.result == "":
                 # This was feedback for query with no result
@@ -141,9 +151,16 @@ class LanceQueryPlanAgent(ChatAgent):
                 return NO_ANSWER
             else:  # non-empty and non-null answer
                 return DONE + " " + self.result
+        self.n_retries += 1
+        if self.n_retries >= self.config.max_retries:
+            # bail out to avoid infinite loop
+            self.n_retries = 0
+            return DONE + " " + NO_ANSWER
         return f"""
-        here is FEEDBACK about your QUERY PLAN. Modify it if needed:
-        {msg.feedback}
+        here is FEEDBACK about your QUERY PLAN, and a SUGGESTED FIX.
+        Modify the QUERY PLAN if needed:
+        FEEDBACK: {msg.feedback}
+        SUGGESTED FIX: {msg.suggested_fix}
         """
 
     def handle_message_fallback(
@@ -177,4 +194,14 @@ class LanceQueryPlanAgent(ChatAgent):
             response_tmpl.metadata.recipient = self.config.critic_name
             self.curr_query_plan = None  # reset
             return response_tmpl
+        if (
+            isinstance(msg, ChatDocument)
+            and not self.has_tool_message_attempt(msg)
+            and msg.metadata.sender == lr.Entity.LLM
+        ):
+            # remind LLM to use the QueryPlanFeedbackTool
+            return """
+            You forgot to use the `query_plan` tool/function.
+            Re-try your response using the `query_plan` tool/function.
+            """
         return None

{langroid-0.1.241 → langroid-0.1.244}/langroid/agent/special/lance_tools.py

@@ -40,5 +40,9 @@ class QueryPlanAnswerTool(ToolMessage):
 
 class QueryPlanFeedbackTool(ToolMessage):
     request = "query_plan_feedback"
-    purpose = "To give <feedback> regarding the query plan."
+    purpose = """
+    To give <feedback> regarding the query plan, 
+    along with a <suggested_fix> if any (empty string if no fix is suggested).
+    """
     feedback: str
+    suggested_fix: str

{langroid-0.1.241 → langroid-0.1.244}/langroid/language_models/openai_gpt.py

@@ -71,6 +71,7 @@ class OpenAIChatModel(str, Enum):
     GPT4 = "gpt-4"
     GPT4_32K = "gpt-4-32k"
     GPT4_TURBO = "gpt-4-turbo"
+    GPT4o = "gpt-4o"
 
 
 class OpenAICompletionModel(str, Enum):
@@ -86,6 +87,7 @@ _context_length: Dict[str, int] = {
     OpenAIChatModel.GPT4: 8192,
     OpenAIChatModel.GPT4_32K: 32_768,
     OpenAIChatModel.GPT4_TURBO: 128_000,
+    OpenAIChatModel.GPT4o: 128_000,
     OpenAICompletionModel.TEXT_DA_VINCI_003: 4096,
 }
 
@@ -95,10 +97,12 @@ _cost_per_1k_tokens: Dict[str, Tuple[float, float]] = {
     OpenAIChatModel.GPT3_5_TURBO: (0.001, 0.002),
     OpenAIChatModel.GPT4: (0.03, 0.06),  # 8K context
     OpenAIChatModel.GPT4_TURBO: (0.01, 0.03),  # 128K context
+    OpenAIChatModel.GPT4o: (0.005, 0.015),  # 128K context
 }
 
 
 openAIChatModelPreferenceList = [
+    OpenAIChatModel.GPT4o,
     OpenAIChatModel.GPT4_TURBO,
     OpenAIChatModel.GPT4,
     OpenAIChatModel.GPT3_5_TURBO,

{langroid-0.1.241 → langroid-0.1.244}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "langroid"
-version = "0.1.241"
+version = "0.1.244"
 description = "Harness LLMs with Multi-Agent Programming"
 authors = ["Prasad Chalasani <pchalasani@gmail.com>"]
 readme = "README.md"
@@ -90,7 +90,7 @@ grpcio = "^1.62.1"
 metaphor-python = {version = "^0.1.23", optional = true}
 chainlit = {version = "^1.0.400", optional = true}
 python-socketio = {version="^5.11.0", optional=true}
-duckduckgo-search = "^5.3.0"
+duckduckgo-search = "^6.0.0"
 huggingface-hub = {version="^0.21.2", optional=true}
 pdf2image = "^1.17.0"
 pytesseract = "^0.3.10"