Added demo to talk to ACL with GPT

pyserini/demo/aclchatgpt/chatbot.py

@@ -0,0 +1,104 @@
+from dataclasses import dataclass
+from argparse import ArgumentParser
+from typing import Optional
+
+import semantic_kernel as sk
+from semantic_kernel.connectors.ai.open_ai import OpenAITextCompletion
+
+from pyserini.demo.aclchatgpt.skill import PyseriniSkill, PyseriniConfig
+
+
+@dataclass
+class OpenAIConfig:
+    api_key: str
+    org_id: str
+
+class ChatBot:
+
+    acl_chat_prompt = """Given the query_results below, your task is to formulate an answer. You may choose to use or not to use the 
+    information in the query_results.
+    If you use the query_results, you must reference the docid of the document used by appending to the answer with "(docid: doc-id-here)".
+    If you do not use the query_results, do not reference it in your answer.
+
+    ===================
+    query_results: {{pyserini.search $input}}
+    ===================
+
+    What is your response to {{$input}}?
+    """
+
+    absolute_question_prompt = """
+    Task: You are an AI language model tasked with transforming given questions into 
+    absolute questions. An absolute question is a question that can stand on its own and carries all the context needed 
+    to be answered. Here's an example:
+
+    User: Who is Alan Turing?
+    ChatBot: Who is Alan Turing?
+    User: How old is he?
+    ChatBot: How old is Alan Turing?
+
+    ===================
+    History: {{$history}}
+    ===================
+
+    Using the history as context, transform the following question into an absolute question: {{$input}}
+    """
+
+    def __init__(self, pyserini_config: PyseriniConfig, openai_config: OpenAIConfig):
+
+        self.kernel = sk.Kernel()
+        self.kernel.add_text_completion_service("dv", OpenAITextCompletion("text-davinci-003", openai_config.api_key, openai_config.org_id))
+        self.kernel.import_skill(PyseriniSkill(pyserini_config),"pyserini")
+        self.context = self.kernel.create_new_context()
+        self.context["url"] = "http://127.0.0.1:8080/search"
+        self.context["history"] = ""
+
+        self.acl_chat_function = self.kernel.create_semantic_function(self.acl_chat_prompt, max_tokens=200, temperature=0, top_p=0.5)
+        self.absolute_question_function = self.kernel.create_semantic_function(self.absolute_question_prompt, max_tokens=200, temperature=0, top_p=0.5)
+
+
+    def _chat(self,input_text: str) -> None:
+
+        print("---------------------------------------------")
+        absolute_question = self.absolute_question_function(input_text,context=self.context)
+
+        print (f"Absolute Question: {absolute_question}")
+
+        # Process the user message and get an answer
+        answer = self.acl_chat_function(str(absolute_question),context=self.context)
+
+        # Show the response
+        print(f"ChatBot: {answer}")
+
+        self.context["history"] += f"\nUser: {input_text}\nChatBot: {answer}\n"
+
+    def chat(self) -> None:
+        print("Hi, I'm the ACL ChatBot. Ask me a question about ACL Anthology papers and I'll do my best to answer it.")
+
+        while True:
+            print("=============================================")
+            self._chat(input("User: "))
+
+def main():
+
+    parser = ArgumentParser()
+
+    parser.add_argument('--k1', type=float, help='BM25 k1 parameter.')
+    parser.add_argument('--b', type=float, help='BM25 b parameter.')
+    parser.add_argument('--hits', type=int, default=10, help='Number of hits returned by the retriever')
+
+    args = parser.parse_args()
+    api_key, org_id = sk.openai_settings_from_dot_env()
+    open_ai_config = OpenAIConfig(api_key,org_id)
+    pyserini_config = PyseriniConfig(args.k1, args.b, args.hits)
+
+    print("Starting ChatBot...")
+
+    chatbot = ChatBot(pyserini_config=pyserini_config,openai_config=open_ai_config)
+    chatbot.chat()
+
+
+
+
+if __name__ == '__main__':
+    main()

pyserini/demo/aclchatgpt/readme.md

@@ -0,0 +1,20 @@
+# Talking to ACL Anthology with GPT
+
+By default, GPT-3 text-davinci-003 is used. You can change this in `chatbot.py`.
+
+## Environment Varaibles
+You will need to create an `.env` file in directory that you are running in.
+The `.env` file should contain the following:
+```
+OPENAI_API_KEY="<your key>"
+OPENAI_ORG_ID="<your org id>"
+```
+
+## Setup
+
+1. Follow the instructions in [Indexing the ACL Anthology with Anserini](https://github.com/castorini/pyserini/blob/master/docs/working-with-acl-anthology.md) to setup the project and generate a `lucene-index-acl-paragraph` index.
+2. Copy the generated `lucene-index-acl-paragraph` index from the `acl-anthology` folder to `pyserini/indexes`
+3. You will need Semantic Kernel as well. 
+   - `pip3 install --upgrade semantic-kernel`
+4. Start the chatbot with `python -m pyserini.demo.aclchatgpt.chatbot`
+5. Start chatting!

pyserini/demo/aclchatgpt/skill.py

@@ -0,0 +1,75 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+import json
+import logging
+from dataclasses import dataclass
+
+import aiohttp
+import requests
+
+from semantic_kernel.orchestration.sk_context import SKContext
+from typing import Callable, Optional, Tuple, Union
+from semantic_kernel.skill_definition import sk_function, sk_function_context_parameter
+from pyserini.search import LuceneSearcher, FaissSearcher, AutoQueryEncoder
+
+@dataclass
+class PyseriniConfig:
+    k1: Optional[float]=None
+    b: Optional[float]=None
+    hits: Optional[int]=1
+
+class PyseriniSkill:
+    """
+    A skill that uses Pyserini to search a corpus of documents.
+
+    Usage:
+        kernel.import_skill(PyseriniSkill(), "http")
+
+    Examples:
+
+        {{pyserini.search $query}}
+    """
+
+    def __init__(self,pyserini_config:PyseriniConfig):
+        self.lang = 'en'
+        self.searcher = LuceneSearcher('indexes/lucene-index-acl-paragraph')
+        self.searcher.set_language(self.lang)
+        if pyserini_config.k1 is not None and pyserini_config.b is not None:
+            self.searcher.set_bm25(pyserini_config.k1, pyserini_config.b)
+            self.retriever_name = f'BM25 (k1={pyserini_config.k1}, b={pyserini_config.b})'
+        else:
+            self.retriever_name = 'BM25'
+        self.hits = pyserini_config.hits
+
+    @sk_function(description="Searches a corpus of documents using Pyserini using the specified query.", name="search")
+    @sk_function_context_parameter(name="url", description="The url of the request")
+    async def search(self, query: str, context: SKContext) -> str:
+        """
+        Searches a corpus of documents using Pyserini using the specified query.
+        Return the response body as a string.
+
+        params:
+            query: The query to search for.
+            context: The SKContext containing the url of the request.
+        returns:
+            The response body as a string.
+        """
+        _, url = context.variables.get("url")
+        if not url:
+            raise ValueError("url cannot be `None` or empty")
+
+        if not query:
+            search_results = []
+        else:
+            hits = self.searcher.search(query, k=self.hits)
+            docs = [self.searcher.doc(hit.docid) for hit in hits]
+            search_results = [
+                {
+                    'rank': r + 1,
+                    'docid': hit.docid,
+                    'doc': docs[r].contents(),
+                }
+                for r, hit in enumerate(hits)
+            ]
+        return "docid:" + search_results[0]["docid"] + ",doc:" + search_results[0]["doc"]
+