Spaces:
Runtime error
Runtime error
parser and prompt template
Browse files
utils.py
CHANGED
|
@@ -1,17 +1,20 @@
|
|
| 1 |
import os
|
| 2 |
import pickle
|
|
|
|
|
|
|
| 3 |
|
| 4 |
-
|
| 5 |
-
from langchain
|
| 6 |
-
from langchain.
|
|
|
|
| 7 |
from langchain.chains import ConversationalRetrievalChain
|
| 8 |
-
from langchain.text_splitter import CharacterTextSplitter
|
| 9 |
from langchain.document_loaders import DirectoryLoader, TextLoader, UnstructuredHTMLLoader
|
| 10 |
-
import faiss
|
| 11 |
-
from langchain.vectorstores.faiss import FAISS
|
| 12 |
from langchain.embeddings import OpenAIEmbeddings
|
| 13 |
-
|
| 14 |
-
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
pickle_file = "open_ai.pkl"
|
| 17 |
index_file = "open_ai.index"
|
|
@@ -26,6 +29,55 @@ memory = ConversationBufferWindowMemory(memory_key="chat_history")
|
|
| 26 |
|
| 27 |
gpt_3_5_index = None
|
| 28 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
def get_search_index():
|
| 30 |
global gpt_3_5_index
|
| 31 |
if os.path.isfile(pickle_file) and os.path.isfile(index_file) and os.path.getsize(pickle_file) > 0:
|
|
|
|
| 1 |
import os
|
| 2 |
import pickle
|
| 3 |
+
import re
|
| 4 |
+
from typing import List, Union
|
| 5 |
|
| 6 |
+
import faiss
|
| 7 |
+
from langchain import OpenAI, LLMChain
|
| 8 |
+
from langchain.agents import ConversationalAgent
|
| 9 |
+
from langchain.agents import Tool, AgentExecutor, LLMSingleActionAgent, AgentOutputParser
|
| 10 |
from langchain.chains import ConversationalRetrievalChain
|
|
|
|
| 11 |
from langchain.document_loaders import DirectoryLoader, TextLoader, UnstructuredHTMLLoader
|
|
|
|
|
|
|
| 12 |
from langchain.embeddings import OpenAIEmbeddings
|
| 13 |
+
from langchain.memory import ConversationBufferWindowMemory
|
| 14 |
+
from langchain.prompts import BaseChatPromptTemplate
|
| 15 |
+
from langchain.schema import AgentAction, AgentFinish, HumanMessage
|
| 16 |
+
from langchain.text_splitter import CharacterTextSplitter
|
| 17 |
+
from langchain.vectorstores.faiss import FAISS
|
| 18 |
|
| 19 |
pickle_file = "open_ai.pkl"
|
| 20 |
index_file = "open_ai.index"
|
|
|
|
| 29 |
|
| 30 |
gpt_3_5_index = None
|
| 31 |
|
| 32 |
+
class CustomOutputParser(AgentOutputParser):
|
| 33 |
+
|
| 34 |
+
def parse(self, llm_output: str) -> Union[AgentAction, AgentFinish]:
|
| 35 |
+
# Check if agent replied without using tools
|
| 36 |
+
if "AI:" in llm_output:
|
| 37 |
+
return AgentFinish(return_values={"output": llm_output.split("AI:")[-1].strip()},
|
| 38 |
+
log=llm_output)
|
| 39 |
+
# Check if agent should finish
|
| 40 |
+
if "Final Answer:" in llm_output:
|
| 41 |
+
return AgentFinish(
|
| 42 |
+
# Return values is generally always a dictionary with a single `output` key
|
| 43 |
+
# It is not recommended to try anything else at the moment :)
|
| 44 |
+
return_values={"output": llm_output.split("Final Answer:")[-1].strip()},
|
| 45 |
+
log=llm_output,
|
| 46 |
+
)
|
| 47 |
+
# Parse out the action and action input
|
| 48 |
+
regex = r"Action: (.*?)[\n]*Action Input:[\s]*(.*)"
|
| 49 |
+
match = re.search(regex, llm_output, re.DOTALL)
|
| 50 |
+
if not match:
|
| 51 |
+
raise ValueError(f"Could not parse LLM output: `{llm_output}`")
|
| 52 |
+
action = match.group(1).strip()
|
| 53 |
+
action_input = match.group(2)
|
| 54 |
+
# Return the action and action input
|
| 55 |
+
return AgentAction(tool=action, tool_input=action_input.strip(" ").strip('"'), log=llm_output)
|
| 56 |
+
|
| 57 |
+
# Set up a prompt template
|
| 58 |
+
class CustomPromptTemplate(BaseChatPromptTemplate):
|
| 59 |
+
# The template to use
|
| 60 |
+
template: str
|
| 61 |
+
# The list of tools available
|
| 62 |
+
tools: List[Tool]
|
| 63 |
+
|
| 64 |
+
def format_messages(self, **kwargs) -> str:
|
| 65 |
+
# Get the intermediate steps (AgentAction, Observation tuples)
|
| 66 |
+
# Format them in a particular way
|
| 67 |
+
intermediate_steps = kwargs.pop("intermediate_steps")
|
| 68 |
+
thoughts = ""
|
| 69 |
+
for action, observation in intermediate_steps:
|
| 70 |
+
thoughts += action.log
|
| 71 |
+
thoughts += f"\nObservation: {observation}\nThought: "
|
| 72 |
+
# Set the agent_scratchpad variable to that value
|
| 73 |
+
kwargs["agent_scratchpad"] = thoughts
|
| 74 |
+
# Create a tools variable from the list of tools provided
|
| 75 |
+
kwargs["tools"] = "\n".join([f"{tool.name}: {tool.description}" for tool in self.tools])
|
| 76 |
+
# Create a list of tool names for the tools provided
|
| 77 |
+
kwargs["tool_names"] = ", ".join([tool.name for tool in self.tools])
|
| 78 |
+
formatted = self.template.format(**kwargs)
|
| 79 |
+
return [HumanMessage(content=formatted)]
|
| 80 |
+
|
| 81 |
def get_search_index():
|
| 82 |
global gpt_3_5_index
|
| 83 |
if os.path.isfile(pickle_file) and os.path.isfile(index_file) and os.path.getsize(pickle_file) > 0:
|