File size: 4,285 Bytes
24ff9b2 2af0eb7 24ff9b2 2af0eb7 24ff9b2 2af0eb7 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 |
# Make sure ollama serve is running(docker or terminal)
# This is the same code as ClearML
from operator import itemgetter
import gradio as gr
from langchain.prompts import PromptTemplate
from shared import getModel, getEmbeddingsModel, getQdrantClient
def answer(samplePrompt, useSample, Query):
if useSample:
query = samplePrompt
else:
query = Query
# Create a qdrant connection
qClient = getQdrantClient()
# Setup the text embedder
embeddingsModel = getEmbeddingsModel()
# Setup the model
model = getModel()
# Retrieval Pipeline
# Retrieve the chunks with the most similar embeddings from Qdrant
def retriever(text, collection):
results = qClient.search(
collection_name=collection,
query_vector = embeddingsModel.embed_query(text),
limit=10
)
return results
# Query expansion(I only generate one additional prompt for simplicity)
template = """
Rewrite the prompt. The new prompt must offer a different perspective.
Do not change the meaning. Output only the rewritten prompt with no introduction.
Prompt: {prompt}
"""
prompt = PromptTemplate.from_template(template)
chain = {"prompt": itemgetter("prompt")} | prompt | model
queryExpansion = chain.invoke({"prompt": query})
# Self-querying(The metadata I will be generating determines whether to look through the Qdrant collection containing github code)
template = """
You are an AI assistant. You must determine if the prompt requires code as the answer.
Output a 1 if it is or a 0 if it is not and nothing else.
Prompt: {prompt}
"""
prompt = PromptTemplate.from_template(template)
chain = {"prompt": itemgetter("prompt")} | prompt | model
codingQuestion = chain.invoke({"prompt": query})
# Filtered vector search for each of the N queries after expansion
relatedCollection = 'Document'
if (codingQuestion == '1'):
relatedCollection = 'Github'
results1 = retriever(query, relatedCollection)
results2 = retriever(queryExpansion, relatedCollection)
# Collecting results
results = results1+results2
# Reranking(Instead of using a CrossEncoder, I will manually compare embeddings)
ids = [result.id for result in results]
scores = [result.score for result in results]
topIds = []
topIndexes = []
for x in range(3):
maxScore = 0
maxIndex = 0
for i in range(len(ids)):
if ids[i] not in topIds and scores[i] > maxScore:
maxScore = scores[i]
maxIndex = i
topIds.append(ids[maxIndex])
topIndexes.append(maxIndex)
texts = [result.payload['text'] for result in results]
links = [result.payload['link'] for result in results]
topTexts = ''
for index in topIndexes:
topTexts += texts[index]
# Building prompt
if(codingQuestion == '1'):
template = """
Write code for the following question given the related coding document below.
Document: {document}
Question: {question}
"""
prompt = PromptTemplate.from_template(template)
else:
template = """
You are an AI agent that has retreived a document from the web.
If the document is useful for answering the question use it.
If the document is not useful, answer normally.
Do not mention the document.
Document: {document}
Question: {question}
"""
prompt = PromptTemplate.from_template(template)
# Obtaining answer
chain = {"document": itemgetter("document"), "question": itemgetter("question")} | prompt | model
return chain.invoke({"document": topTexts, "question": query})
demo = gr.Interface(
fn=answer,
inputs=[
gr.Dropdown(
["How can I develop the navigation stack of an agent with egomotion?",
"What is ROS?", "How many companies is Nav2 trusted by worldwide?",
"How would I build a ROS 2 Navigation Framework and System?",
"Write me code to move a robot using Moveit"], label="Sample Prompt"
),
"checkbox",
"text",
],
outputs=["text"],
)
demo.launch(share=False)
|