Commit
·
c8febd3
1
Parent(s):
a276d30
fix: try again
Browse files
app.py
CHANGED
@@ -2,12 +2,20 @@ import os
|
|
2 |
import gradio as gr
|
3 |
import requests
|
4 |
import inspect
|
|
|
5 |
from llama_index.llms.huggingface_api import HuggingFaceInferenceAPI
|
6 |
from llama_index.readers.youtube_transcript import YoutubeTranscriptReader
|
7 |
from llama_index.core.tools import FunctionTool
|
8 |
from llama_index.tools.duckduckgo import DuckDuckGoSearchToolSpec
|
9 |
from llama_index.core.agent.workflow import AgentWorkflow
|
10 |
from llama_index.llms.gemini import Gemini
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
import pandas as pd
|
12 |
import asyncio
|
13 |
|
@@ -19,15 +27,15 @@ DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
|
19 |
# ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
|
20 |
class BasicAgent:
|
21 |
def __init__(self):
|
22 |
-
self.llm =
|
|
|
23 |
print("BasicAgent initialized.")
|
24 |
def __call__(self, question: str) -> str:
|
25 |
print(f"Agent received question (first 50 chars): {question[:50]}...")
|
26 |
-
def
|
27 |
"""Fetches transcript of the given youtube_link and returns matching answers based on query.
|
28 |
Args:
|
29 |
youtube_link (str): youtube video link for which we need to answer questions on.
|
30 |
-
query (str): question to answer from video transcript.
|
31 |
"""
|
32 |
loader = YoutubeTranscriptReader()
|
33 |
documents = loader.load_data(
|
@@ -39,21 +47,46 @@ class BasicAgent:
|
|
39 |
return "Here's the transcript from the video, examine and formulate answer based on what is said in the transcript: \n" + text
|
40 |
|
41 |
youtube_transcript_answer_tool = FunctionTool.from_defaults(
|
42 |
-
|
43 |
-
name="
|
44 |
-
description="Fetches transcript of the given
|
45 |
)
|
46 |
|
47 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
|
49 |
-
agent = AgentWorkflow.from_tools_or_functions([
|
50 |
|
51 |
async def run_agent():
|
52 |
return await agent.run(question)
|
53 |
|
54 |
-
|
55 |
-
|
56 |
-
|
|
|
57 |
|
58 |
def run_and_submit_all( profile: gr.OAuthProfile | None):
|
59 |
"""
|
|
|
2 |
import gradio as gr
|
3 |
import requests
|
4 |
import inspect
|
5 |
+
from llama_index.llms.ollama import Ollama
|
6 |
from llama_index.llms.huggingface_api import HuggingFaceInferenceAPI
|
7 |
from llama_index.readers.youtube_transcript import YoutubeTranscriptReader
|
8 |
from llama_index.core.tools import FunctionTool
|
9 |
from llama_index.tools.duckduckgo import DuckDuckGoSearchToolSpec
|
10 |
from llama_index.core.agent.workflow import AgentWorkflow
|
11 |
from llama_index.llms.gemini import Gemini
|
12 |
+
from llama_index.retrievers.bm25 import BM25Retriever
|
13 |
+
from llama_index.core.query_engine import RetrieverQueryEngine
|
14 |
+
from llama_index.core.tools import QueryEngineTool
|
15 |
+
from llama_index.core.node_parser import SentenceSplitter
|
16 |
+
from llama_index.llms.huggingface_api import HuggingFaceInferenceAPI
|
17 |
+
from llama_index.core.schema import Document
|
18 |
+
from llama_index.core import get_response_synthesizer
|
19 |
import pandas as pd
|
20 |
import asyncio
|
21 |
|
|
|
27 |
# ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
|
28 |
class BasicAgent:
|
29 |
def __init__(self):
|
30 |
+
# self.llm = Ollama(model="qwen2.5:7b", request_timeout=500)
|
31 |
+
self.llm = Gemini(model_name="models/gemini-2.0-flash-lite")
|
32 |
print("BasicAgent initialized.")
|
33 |
def __call__(self, question: str) -> str:
|
34 |
print(f"Agent received question (first 50 chars): {question[:50]}...")
|
35 |
+
def get_youtube_transcript(youtube_link: str) -> str:
|
36 |
"""Fetches transcript of the given youtube_link and returns matching answers based on query.
|
37 |
Args:
|
38 |
youtube_link (str): youtube video link for which we need to answer questions on.
|
|
|
39 |
"""
|
40 |
loader = YoutubeTranscriptReader()
|
41 |
documents = loader.load_data(
|
|
|
47 |
return "Here's the transcript from the video, examine and formulate answer based on what is said in the transcript: \n" + text
|
48 |
|
49 |
youtube_transcript_answer_tool = FunctionTool.from_defaults(
|
50 |
+
get_youtube_transcript,
|
51 |
+
name="get_youtube_transcript",
|
52 |
+
description="Fetches transcript of the given video using youtube_link.",
|
53 |
)
|
54 |
|
55 |
+
def duck_duck_go_search_tool(query: str) -> str:
|
56 |
+
try:
|
57 |
+
raw_results = DuckDuckGoSearchToolSpec().duckduckgo_full_search(query)
|
58 |
+
texts = [res['body'] for res in raw_results]
|
59 |
+
documents = [Document(text=body) for body in texts]
|
60 |
+
|
61 |
+
splitter = SentenceSplitter(chunk_size=256)
|
62 |
+
nodes = splitter.get_nodes_from_documents(documents)
|
63 |
+
|
64 |
+
retriever = BM25Retriever(nodes=nodes, similarity_top_k=10)
|
65 |
+
synthesizer = get_response_synthesizer(response_mode="refine", llm=self.llm)
|
66 |
+
query_engine = RetrieverQueryEngine(retriever=retriever, response_synthesizer=synthesizer)
|
67 |
+
|
68 |
+
response = query_engine.query(query)
|
69 |
+
return response.response
|
70 |
+
|
71 |
+
except Exception as e:
|
72 |
+
return f"An error occurred: {e}"
|
73 |
+
|
74 |
+
|
75 |
+
duckduckgo_search_tool = FunctionTool.from_defaults(
|
76 |
+
duck_duck_go_search_tool,
|
77 |
+
name="duck_duck_go_search_tool",
|
78 |
+
description="Searches the web and refines the result into a high-quality answer."
|
79 |
+
)
|
80 |
|
81 |
+
agent = AgentWorkflow.from_tools_or_functions([duckduckgo_search_tool, youtube_transcript_answer_tool], llm=self.llm)
|
82 |
|
83 |
async def run_agent():
|
84 |
return await agent.run(question)
|
85 |
|
86 |
+
response = asyncio.run(run_agent())
|
87 |
+
final_answer = response.response.blocks[0].text
|
88 |
+
print(f"Agent returning fixed answer: {final_answer}")
|
89 |
+
return final_answer
|
90 |
|
91 |
def run_and_submit_all( profile: gr.OAuthProfile | None):
|
92 |
"""
|