stash

docs/examples/demos/react.ipynb

@@ -68,7 +68,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "id": "b5a265c6",
+   "id": "2d6f4b1d",
    "metadata": {},
    "outputs": [],
    "source": []

docs/examples/demos/self_ask_with_search.ipynb

@@ -20,23 +20,28 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
+      "\n",
+      "\n",
+      "\u001b[1m> Entering new chain...\u001b[0m\n",
       "\n",
       "\n",
       "\u001b[1m> Entering new chain...\u001b[0m\n",
       "What is the hometown of the reigning men's U.S. Open champion?\n",
       "Are follow up questions needed here:\u001b[32;1m\u001b[1;3m Yes.\n",
       "Follow up: Who is the reigning men's U.S. Open champion?\u001b[0m\n",
-      "Intermediate answer: \u001b[33;1m\u001b[1;3mCarlos Alcaraz.\u001b[0m\u001b[32;1m\u001b[1;3m\n",
-      "Follow up: Where is Carlos Alcaraz from?\u001b[0m\n",
-      "Intermediate answer: \u001b[33;1m\u001b[1;3mEl Palmar, Spain.\u001b[0m\u001b[32;1m\u001b[1;3m\n",
-      "So the final answer is: El Palmar, Spain\u001b[0m\n",
+      "Intermediate answer: \u001b[36;1m\u001b[1;3mCarlos Alcaraz\u001b[0m\n",
+      "\u001b[32;1m\u001b[1;3mFollow up: Where is Carlos Alcaraz from?\u001b[0m\n",
+      "Intermediate answer: \u001b[36;1m\u001b[1;3mEl Palmar, Spain\u001b[0m\n",
+      "\u001b[32;1m\u001b[1;3mSo the final answer is: El Palmar, Spain\u001b[0m\n",
+      "\u001b[1m> Finished chain.\u001b[0m\n",
+      "\n",
       "\u001b[1m> Finished chain.\u001b[0m\n"
      ]
     },
     {
      "data": {
       "text/plain": [
-       "'\\nSo the final answer is: El Palmar, Spain'"
+       "'El Palmar, Spain'"
       ]
      },
      "execution_count": 1,

langchain/chains/mrkl/base.py

@@ -10,6 +10,7 @@ from langchain.chains.router import LLMRouterChain
 from langchain.input import ChainedInput, get_color_mapping
 from langchain.llms.base import LLM
 from langchain.prompts import BasePromptTemplate, PromptTemplate
+from langchain.chains.router_expert import RouterExpertChain, ExpertConfig
 
 FINAL_ANSWER_ACTION = "Final Answer: "
 
@@ -31,9 +32,9 @@ class ChainConfig(NamedTuple):
 def get_action_and_input(llm_output: str) -> Tuple[str, str]:
     """Parse out the action and input from the LLM output."""
     ps = [p for p in llm_output.split("\n") if p]
-    if ps[-1].startswith(FINAL_ANSWER_ACTION):
+    if ps[-1].startswith("Final Answer"):
         directive = ps[-1][len(FINAL_ANSWER_ACTION) :]
-        return FINAL_ANSWER_ACTION, directive
+        return "Final Answer", directive
     if not ps[-1].startswith("Action Input: "):
         raise ValueError(
             "The last line does not have an action input, "
@@ -52,6 +53,16 @@ def get_action_and_input(llm_output: str) -> Tuple[str, str]:
 class MRKLRouterChain(LLMRouterChain):
     """Router for the MRKL chain."""
 
+    @property
+    def observation_prefix(self) -> str:
+        """Prefix to append the observation with."""
+        return "Observation: "
+
+    @property
+    def router_prefix(self) -> str:
+        """Prefix to append the router call with."""
+        return "Thought:"
+
     def __init__(self, llm: LLM, chain_configs: List[ChainConfig], **kwargs: Any):
         """Initialize with an LLM and the chain configs it has access to."""
         tools = "\n".join(
@@ -166,21 +177,15 @@ class MRKLChain(Chain, BaseModel):
 
     def _call(self, inputs: Dict[str, str]) -> Dict[str, str]:
         router_chain = MRKLRouterChain(self.llm, self.chain_configs)
-        chained_input = ChainedInput(
-            f"{inputs[self.input_key]}\nThought:", verbose=self.verbose
-        )
-        color_mapping = get_color_mapping(
-            list(self.action_to_chain_map.keys()), excluded_colors=["green"]
+        question = inputs[self.input_key]
+        expert_configs = [
+            ExpertConfig(expert_name=c.action_name, expert=c.action)
+            for c in self.chain_configs
+        ]
+        chain = RouterExpertChain(
+            router_chain=router_chain,
+            expert_configs=expert_configs,
+            verbose=self.verbose
         )
-        while True:
-            action, action_input, thought = router_chain.get_action_and_input(
-                chained_input.input
-            )
-            chained_input.add(thought, color="green")
-            if action == FINAL_ANSWER_ACTION:
-                return {self.output_key: action_input}
-            chain = self.action_to_chain_map[action]
-            ca = chain(action_input)
-            chained_input.add("\nObservation: ")
-            chained_input.add(ca, color=color_mapping[action])
-            chained_input.add("\nThought:")
+        output = chain.run(question)
+        return {self.output_key: output}

langchain/chains/react/base.py

@@ -12,6 +12,7 @@ from langchain.docstore.base import Docstore
 from langchain.docstore.document import Document
 from langchain.input import ChainedInput
 from langchain.llms.base import LLM
+from langchain.chains.router_expert import RouterExpertChain, ExpertConfig
 
 
 class ReActRouterChain(LLMRouterChain, BaseModel):
@@ -43,6 +44,42 @@ class ReActRouterChain(LLMRouterChain, BaseModel):
             raise ValueError(f"Could not parse action directive: {action_str}")
         return re_matches.group(1), re_matches.group(2)
 
+    @property
+    def finish_action_name(self) -> str:
+        """The action name of when to finish the chain."""
+        return "Finish"
+
+    @property
+    def observation_prefix(self) -> str:
+        """Prefix to append the observation with."""
+        return f"Observation {self.i - 1}: "
+
+    @property
+    def router_prefix(self) -> str:
+        """Prefix to append the router call with."""
+        return f"Thought {self.i}:"
+
+
+class DocstoreExplorer:
+
+    def __init__(self, docstore: Docstore):
+        self.docstore=docstore
+        self.document = None
+
+    def search(self, term: str):
+        result = self.docstore.search(term)
+        if isinstance(result, Document):
+            self.document = result
+            return self.document.summary
+        else:
+            self.document = None
+            return result
+
+    def lookup(self, term: str):
+        if self.document is None:
+            raise ValueError("Cannot lookup without a successful search first")
+        return self.document.lookup(term)
+
 
 class ReActChain(Chain, BaseModel):
     """Chain that implements the ReAct paper.
@@ -86,29 +123,15 @@ class ReActChain(Chain, BaseModel):
     def _call(self, inputs: Dict[str, Any]) -> Dict[str, str]:
         question = inputs[self.input_key]
         router_chain = ReActRouterChain(self.llm)
-        chained_input = ChainedInput(f"{question}\nThought 1:", verbose=self.verbose)
-        document = None
-        while True:
-            action, directive, ret_text = router_chain.get_action_and_input(
-                chained_input.input
-            )
-            chained_input.add(ret_text, color="green")
-            if action == "Search":
-                result = self.docstore.search(directive)
-                if isinstance(result, Document):
-                    document = result
-                    observation = document.summary
-                else:
-                    document = None
-                    observation = result
-            elif action == "Lookup":
-                if document is None:
-                    raise ValueError("Cannot lookup without a successful search first")
-                observation = document.lookup(directive)
-            elif action == "Finish":
-                return {self.output_key: directive}
-            else:
-                raise ValueError(f"Got unknown action directive: {action}")
-            chained_input.add(f"\nObservation {router_chain.i - 1}: ")
-            chained_input.add(observation, color="yellow")
-            chained_input.add(f"\nThought {router_chain.i}:")
+        docstore_explorer = DocstoreExplorer(self.docstore)
+        expert_configs = [
+            ExpertConfig(expert_name="Search", expert=docstore_explorer.search),
+            ExpertConfig(expert_name="Lookup", expert=docstore_explorer.lookup)
+        ]
+        chain = RouterExpertChain(
+            router_chain=router_chain,
+            expert_configs=expert_configs,
+            verbose=self.verbose
+        )
+        output = chain.run(question)
+        return {self.output_key: output}

langchain/chains/router.py

@@ -36,6 +36,21 @@ class RouterChain(Chain, BaseModel, ABC):
     def get_action_and_input(self, text: str) -> Tuple[str, str, str]:
         """Return action, action input, and log (in that order)."""
 
+    @property
+    @abstractmethod
+    def observation_prefix(self) -> str:
+        """Prefix to append the observation with."""
+
+    @property
+    @abstractmethod
+    def router_prefix(self) -> str:
+        """Prefix to append the router call with."""
+
+    @property
+    def finish_action_name(self) -> str:
+        """The action name of when to finish the chain."""
+        return "Final Answer"
+
     def _call(self, inputs: Dict[str, str]) -> Dict[str, str]:
         action, action_input, log = self.get_action_and_input(inputs[self.input_key])
         return {

langchain/chains/router_expert.py

@@ -12,7 +12,7 @@ from langchain.llms.base import LLM
 from langchain.prompts import BasePromptTemplate, PromptTemplate
 from langchain.chains.router import RouterChain
 
-FINAL_ANSWER_ACTION = "Final Answer: "
+
 
 
 class ExpertConfig(NamedTuple):
@@ -28,6 +28,8 @@ class RouterExpertChain(Chain, BaseModel):
     """Router chain."""
     expert_configs: List[ExpertConfig]
     """Expert configs this chain has access to."""
+    starter_string: str = "\n"
+    """String to put after user input but before first router."""
     input_key: str = "question"  #: :meta private:
     output_key: str = "answer"  #: :meta private:
 
@@ -54,22 +56,22 @@ class RouterExpertChain(Chain, BaseModel):
         return [self.output_key]
 
     def _call(self, inputs: Dict[str, str]) -> Dict[str, str]:
-        router_chain = MRKLRouterChain(self.llm, self.chain_configs)
+        action_to_chain_map = {e.expert_name: e.expert for e in self.expert_configs}
         chained_input = ChainedInput(
-            f"{inputs[self.input_key]}", verbose=self.verbose
+            f"{inputs[self.input_key]}{self.starter_string}{self.router_chain.router_prefix}", verbose=self.verbose
         )
         color_mapping = get_color_mapping(
-            [c.], excluded_colors=["green"]
+            [c.expert_name for c in self.expert_configs], excluded_colors=["green"]
         )
         while True:
-            action, action_input, thought = router_chain.get_action_and_input(
+            action, action_input, log = self.router_chain.get_action_and_input(
                 chained_input.input
             )
-            chained_input.add(thought, color="green")
-            if action == FINAL_ANSWER_ACTION:
+            chained_input.add(log, color="green")
+            if action == self.router_chain.finish_action_name:
                 return {self.output_key: action_input}
-            chain = self.action_to_chain_map[action]
+            chain = action_to_chain_map[action]
             ca = chain(action_input)
-            chained_input.add("\nObservation: ")
+            chained_input.add(f"\n{self.router_chain.observation_prefix}")
             chained_input.add(ca, color=color_mapping[action])
-            chained_input.add("\nThought:")
+            chained_input.add(f"\n{self.router_chain.router_prefix}")

langchain/chains/self_ask_with_search/base.py

@@ -10,6 +10,7 @@ from langchain.chains.self_ask_with_search.prompt import PROMPT
 from langchain.chains.serpapi import SerpAPIChain
 from langchain.input import ChainedInput
 from langchain.llms.base import LLM
+from langchain.chains.router_expert import RouterExpertChain, ExpertConfig
 
 
 class SelfAskWithSearchRouter(LLMRouterChain):
@@ -28,7 +29,10 @@ class SelfAskWithSearchRouter(LLMRouterChain):
             last_line = text.split("\n")[-1]
 
         if followup not in last_line:
-            return "Final Answer", text
+            finish_string = "So the final answer is: "
+            if finish_string not in last_line:
+                raise ValueError("We should probably never get here")
+            return "Final Answer", text[len(finish_string):]
 
         if ":" not in last_line:
             after_colon = last_line
@@ -42,6 +46,16 @@ class SelfAskWithSearchRouter(LLMRouterChain):
 
         return "Intermediate Answer", after_colon
 
+    @property
+    def observation_prefix(self) -> str:
+        """Prefix to append the observation with."""
+        return "Intermediate answer: "
+
+    @property
+    def router_prefix(self) -> str:
+        """Prefix to append the router call with."""
+        return ""
+
 
 class SelfAskWithSearchChain(Chain, BaseModel):
     """Chain that does self ask with search.
@@ -84,16 +98,9 @@ class SelfAskWithSearchChain(Chain, BaseModel):
         return [self.output_key]
 
     def _call(self, inputs: Dict[str, Any]) -> Dict[str, str]:
-        chained_input = ChainedInput(inputs[self.input_key], verbose=self.verbose)
-        chained_input.add("\nAre follow up questions needed here:")
         intermediate = "\nIntermediate answer:"
         router = SelfAskWithSearchRouter(self.llm, stops=[intermediate])
-        action, action_input, log = router.get_action_and_input(chained_input.input)
-        chained_input.add(log, color="green")
-        while action != "Final Answer":
-            external_answer = self.search_chain.run(action_input)
-            chained_input.add(intermediate + " ")
-            chained_input.add(external_answer + ".", color="yellow")
-            action, action_input, log = router.get_action_and_input(chained_input.input)
-            chained_input.add(log, color="green")
-        return {self.output_key: action_input}
+        expert_configs = [ExpertConfig(expert_name="Intermediate Answer", expert=self.search_chain.run)]
+        chain = RouterExpertChain(router_chain=router, expert_configs=expert_configs, verbose=self.verbose, starter_string="\nAre follow up questions needed here:")
+        output = chain.run(inputs[self.input_key])
+        return {self.output_key: output}