Spaces:
Sleeping
Sleeping
mriusero
commited on
Commit
·
3bcd8f6
1
Parent(s):
439e39d
feat: retrieval (1st version)
Browse files- .gitignore +6 -2
- prompt.md +1 -1
- src/inference.py +3 -0
- src/tools/__init__.py +2 -1
- src/tools/retrieve_knowledge.py +22 -0
- src/tools/visit_webpage.py +11 -3
- src/utils/vector_store.py +124 -0
- tools.json +27 -0
.gitignore
CHANGED
@@ -11,9 +11,13 @@ my-traffic-analysis-441217-32bda1474a0f.json
|
|
11 |
# Python
|
12 |
*__pycache__/
|
13 |
|
14 |
-
#
|
15 |
llm/
|
16 |
attachments/
|
17 |
logs/
|
18 |
1st_run/
|
19 |
-
metadata.jsonl
|
|
|
|
|
|
|
|
|
|
11 |
# Python
|
12 |
*__pycache__/
|
13 |
|
14 |
+
# Project
|
15 |
llm/
|
16 |
attachments/
|
17 |
logs/
|
18 |
1st_run/
|
19 |
+
metadata.jsonl
|
20 |
+
tests.py
|
21 |
+
|
22 |
+
chroma_db/
|
23 |
+
*.bin
|
prompt.md
CHANGED
@@ -2,7 +2,7 @@ You are a general and precise AI assistant. I will ask you a question.
|
|
2 |
Report your thoughts, and finish
|
3 |
your answer with the following template: FINAL ANSWER: [YOUR FINAL ANSWER].
|
4 |
If a tool provide an error, use the tool differently.
|
5 |
-
For web searching, ensure your answer by cross-checking data with several sources.
|
6 |
YOUR FINAL ANSWER should be a number OR as few words as possible OR a comma separated list of
|
7 |
numbers and/or strings.
|
8 |
If you are asked for a number, don’t use comma to write your number neither use units such as $ or percent
|
|
|
2 |
Report your thoughts, and finish
|
3 |
your answer with the following template: FINAL ANSWER: [YOUR FINAL ANSWER].
|
4 |
If a tool provide an error, use the tool differently.
|
5 |
+
For web searching, first search in your knowledge and if necessary complete them with web_search and ensure your answer by cross-checking data with several sources.
|
6 |
YOUR FINAL ANSWER should be a number OR as few words as possible OR a comma separated list of
|
7 |
numbers and/or strings.
|
8 |
If you are asked for a number, don’t use comma to write your number neither use units such as $ or percent
|
src/inference.py
CHANGED
@@ -19,6 +19,7 @@ from src.tools import (
|
|
19 |
analyze_excel,
|
20 |
analyze_youtube_video,
|
21 |
calculate_sum,
|
|
|
22 |
)
|
23 |
|
24 |
load_dotenv()
|
@@ -43,6 +44,7 @@ class Agent:
|
|
43 |
"analyze_excel": analyze_excel,
|
44 |
"analyze_youtube_video": analyze_youtube_video,
|
45 |
"calculate_sum": calculate_sum,
|
|
|
46 |
}
|
47 |
self.log = []
|
48 |
self.tools = self.get_tools()
|
@@ -74,6 +76,7 @@ class Agent:
|
|
74 |
analyze_excel,
|
75 |
analyze_youtube_video,
|
76 |
calculate_sum,
|
|
|
77 |
]
|
78 |
).get('tools')
|
79 |
|
|
|
19 |
analyze_excel,
|
20 |
analyze_youtube_video,
|
21 |
calculate_sum,
|
22 |
+
retrieve_knowledge,
|
23 |
)
|
24 |
|
25 |
load_dotenv()
|
|
|
44 |
"analyze_excel": analyze_excel,
|
45 |
"analyze_youtube_video": analyze_youtube_video,
|
46 |
"calculate_sum": calculate_sum,
|
47 |
+
"retrieve_knowledge": retrieve_knowledge,
|
48 |
}
|
49 |
self.log = []
|
50 |
self.tools = self.get_tools()
|
|
|
76 |
analyze_excel,
|
77 |
analyze_youtube_video,
|
78 |
calculate_sum,
|
79 |
+
retrieve_knowledge,
|
80 |
]
|
81 |
).get('tools')
|
82 |
|
src/tools/__init__.py
CHANGED
@@ -9,4 +9,5 @@ from .transcribe_audio import transcribe_audio
|
|
9 |
from .execute_code import execute_code
|
10 |
from .analyze_excel import analyze_excel
|
11 |
from .analyze_youtube_video import analyze_youtube_video
|
12 |
-
from .calculator import calculate_sum
|
|
|
|
9 |
from .execute_code import execute_code
|
10 |
from .analyze_excel import analyze_excel
|
11 |
from .analyze_youtube_video import analyze_youtube_video
|
12 |
+
from .calculator import calculate_sum
|
13 |
+
from .retrieve_knowledge import retrieve_knowledge
|
src/tools/retrieve_knowledge.py
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from src.utils.tooling import tool
|
2 |
+
|
3 |
+
@tool
|
4 |
+
def retrieve_knowledge(query: str, n_results: int = 5, distance_threshold : float = 0.5) -> str:
|
5 |
+
"""
|
6 |
+
Retrieves knowledge from a database with a provided query.
|
7 |
+
Args:
|
8 |
+
query (str): The query to search for in the vector store.
|
9 |
+
n_results (int, optional): The number of results to return. Default is 5.
|
10 |
+
distance_threshold (float, optional): The minimum distance score for results. Default is 0.5.
|
11 |
+
"""
|
12 |
+
try:
|
13 |
+
from src.utils.vector_store import retrieve_from_database
|
14 |
+
results = retrieve_from_database(
|
15 |
+
query=query,
|
16 |
+
n_results=n_results,
|
17 |
+
distance_threshold=distance_threshold
|
18 |
+
)
|
19 |
+
return str(results)
|
20 |
+
|
21 |
+
except Exception as e:
|
22 |
+
return f"An unexpected error occurred: {str(e)}"
|
src/tools/visit_webpage.py
CHANGED
@@ -1,6 +1,5 @@
|
|
1 |
-
import re
|
2 |
-
|
3 |
from src.utils.tooling import tool
|
|
|
4 |
|
5 |
@tool
|
6 |
def visit_webpage(url: str) -> str:
|
@@ -11,6 +10,7 @@ def visit_webpage(url: str) -> str:
|
|
11 |
url (str): The URL of the webpage to visit.
|
12 |
"""
|
13 |
try:
|
|
|
14 |
import requests
|
15 |
from markdownify import markdownify
|
16 |
from requests.exceptions import RequestException
|
@@ -28,6 +28,14 @@ def visit_webpage(url: str) -> str:
|
|
28 |
markdown_content = markdownify(response.text).strip() # Convert the HTML content to Markdown
|
29 |
markdown_content = re.sub(r"\n{3,}", "\n\n", markdown_content) # Remove multiple line breaks
|
30 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
return truncate_content(markdown_content, 10000)
|
32 |
|
33 |
except requests.exceptions.Timeout:
|
@@ -37,4 +45,4 @@ def visit_webpage(url: str) -> str:
|
|
37 |
return f"Error fetching the webpage: {str(e)}"
|
38 |
|
39 |
except Exception as e:
|
40 |
-
return f"An unexpected error occurred: {str(e)}"
|
|
|
|
|
|
|
1 |
from src.utils.tooling import tool
|
2 |
+
from src.utils.vector_store import vectorize, load_in_vector_db
|
3 |
|
4 |
@tool
|
5 |
def visit_webpage(url: str) -> str:
|
|
|
10 |
url (str): The URL of the webpage to visit.
|
11 |
"""
|
12 |
try:
|
13 |
+
import re
|
14 |
import requests
|
15 |
from markdownify import markdownify
|
16 |
from requests.exceptions import RequestException
|
|
|
28 |
markdown_content = markdownify(response.text).strip() # Convert the HTML content to Markdown
|
29 |
markdown_content = re.sub(r"\n{3,}", "\n\n", markdown_content) # Remove multiple line breaks
|
30 |
|
31 |
+
# Adding metadata
|
32 |
+
metadatas = {
|
33 |
+
"url": url,
|
34 |
+
}
|
35 |
+
|
36 |
+
text_embeddings, chunks = vectorize(markdown_content) # Vectorize the content
|
37 |
+
load_in_vector_db(text_embeddings, chunks, metadatas=metadatas) # Load the text embeddings into a FAISS index
|
38 |
+
|
39 |
return truncate_content(markdown_content, 10000)
|
40 |
|
41 |
except requests.exceptions.Timeout:
|
|
|
45 |
return f"Error fetching the webpage: {str(e)}"
|
46 |
|
47 |
except Exception as e:
|
48 |
+
return f"An unexpected error occurred: {str(e)}"
|
src/utils/vector_store.py
ADDED
@@ -0,0 +1,124 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from dotenv import load_dotenv
|
3 |
+
from mistralai import Mistral
|
4 |
+
import numpy as np
|
5 |
+
import time
|
6 |
+
import chromadb
|
7 |
+
import json
|
8 |
+
|
9 |
+
load_dotenv()
|
10 |
+
MISTRAL_API_KEY = os.getenv("MISTRAL_API_KEY")
|
11 |
+
COLLECTION_NAME = "webpages_collection"
|
12 |
+
PERSIST_DIRECTORY = "./chroma_db"
|
13 |
+
|
14 |
+
def get_text_embeddings(input_texts):
|
15 |
+
"""
|
16 |
+
Get the text embeddings for the given inputs using Mistral API.
|
17 |
+
"""
|
18 |
+
client = Mistral(api_key=MISTRAL_API_KEY)
|
19 |
+
while True:
|
20 |
+
try:
|
21 |
+
embeddings_batch_response = client.embeddings.create(
|
22 |
+
model="mistral-embed",
|
23 |
+
inputs=input_texts
|
24 |
+
)
|
25 |
+
return [data.embedding for data in embeddings_batch_response.data]
|
26 |
+
except Exception as e:
|
27 |
+
if "rate limit exceeded" in str(e).lower():
|
28 |
+
print("Rate limit exceeded. Retrying after 1 second...")
|
29 |
+
time.sleep(1)
|
30 |
+
else:
|
31 |
+
raise
|
32 |
+
|
33 |
+
def vectorize(markdown_content, chunk_size=2048):
|
34 |
+
"""
|
35 |
+
Vectorizes the given markdown content into chunks of specified size.
|
36 |
+
"""
|
37 |
+
chunks = [markdown_content[i:i + chunk_size] for i in range(0, len(markdown_content), chunk_size)]
|
38 |
+
text_embeddings = get_text_embeddings(chunks)
|
39 |
+
return np.array(text_embeddings), chunks
|
40 |
+
|
41 |
+
def load_in_vector_db(text_embeddings, chunks, metadatas=None, collection_name=COLLECTION_NAME):
|
42 |
+
"""
|
43 |
+
Load the text embeddings into a ChromaDB collection for efficient similarity search.
|
44 |
+
"""
|
45 |
+
client = chromadb.PersistentClient(path=PERSIST_DIRECTORY)
|
46 |
+
|
47 |
+
# Check if the collection exists, if not, create it
|
48 |
+
if collection_name not in [col.name for col in client.list_collections()]:
|
49 |
+
collection = client.create_collection(collection_name)
|
50 |
+
else:
|
51 |
+
collection = client.get_collection(collection_name)
|
52 |
+
|
53 |
+
for embedding, chunk in zip(text_embeddings, chunks):
|
54 |
+
collection.add(
|
55 |
+
embeddings=[embedding],
|
56 |
+
documents=[chunk],
|
57 |
+
metadatas=[metadatas],
|
58 |
+
ids=[str(hash(chunk))]
|
59 |
+
)
|
60 |
+
|
61 |
+
|
62 |
+
def see_database(collection_name=COLLECTION_NAME):
|
63 |
+
"""
|
64 |
+
Load the ChromaDB collection and text chunks.
|
65 |
+
"""
|
66 |
+
client = chromadb.PersistentClient(path=PERSIST_DIRECTORY)
|
67 |
+
|
68 |
+
if collection_name not in [col.name for col in client.list_collections()]:
|
69 |
+
print("Collection not found. Please ensure it is created.")
|
70 |
+
return
|
71 |
+
|
72 |
+
collection = client.get_collection(collection_name)
|
73 |
+
|
74 |
+
items = collection.get()
|
75 |
+
|
76 |
+
print(f"Type of items: {type(items)}")
|
77 |
+
print(f"Items: {items}")
|
78 |
+
|
79 |
+
for item in items:
|
80 |
+
print(f"Type of item: {type(item)}")
|
81 |
+
print(f"Item: {item}")
|
82 |
+
|
83 |
+
if isinstance(item, dict):
|
84 |
+
print(f"ID: {item.get('ids')}")
|
85 |
+
print(f"Document: {item.get('document')}")
|
86 |
+
print(f"Metadata: {item.get('metadata')}")
|
87 |
+
else:
|
88 |
+
print("Item is not a dictionary")
|
89 |
+
|
90 |
+
print("---")
|
91 |
+
|
92 |
+
def retrieve_from_database(query, collection_name=COLLECTION_NAME, n_results=5, distance_threshold=None):
|
93 |
+
"""
|
94 |
+
Retrieve the most similar documents from the vector store based on the query.
|
95 |
+
"""
|
96 |
+
client = chromadb.PersistentClient(path=PERSIST_DIRECTORY)
|
97 |
+
collection = client.get_collection(collection_name)
|
98 |
+
query_embeddings = get_text_embeddings([query])
|
99 |
+
raw_results = collection.query(
|
100 |
+
query_embeddings=query_embeddings,
|
101 |
+
n_results=n_results,
|
102 |
+
include=["documents", "metadatas", "distances"]
|
103 |
+
)
|
104 |
+
if distance_threshold is not None:
|
105 |
+
filtered_results = {
|
106 |
+
"ids": [],
|
107 |
+
"distances": [],
|
108 |
+
"metadatas": [],
|
109 |
+
"documents": []
|
110 |
+
}
|
111 |
+
for i, distance in enumerate(raw_results['distances'][0]):
|
112 |
+
if distance >= distance_threshold:
|
113 |
+
filtered_results['ids'].append(raw_results['ids'][0][i])
|
114 |
+
filtered_results['distances'].append(distance)
|
115 |
+
filtered_results['metadatas'].append(raw_results['metadatas'][0][i])
|
116 |
+
filtered_results['documents'].append(raw_results['documents'][0][i])
|
117 |
+
results = filtered_results
|
118 |
+
|
119 |
+
if len(results['documents']) == 0:
|
120 |
+
return "No relevant data found in knowledge database, have you visited webpages?"
|
121 |
+
else:
|
122 |
+
return json.dumps(results, indent=4)
|
123 |
+
else:
|
124 |
+
return json.dumps(raw_results, indent=4)
|
tools.json
CHANGED
@@ -253,5 +253,32 @@
|
|
253 |
]
|
254 |
}
|
255 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
256 |
}
|
257 |
]
|
|
|
253 |
]
|
254 |
}
|
255 |
}
|
256 |
+
},
|
257 |
+
{
|
258 |
+
"type": "function",
|
259 |
+
"function": {
|
260 |
+
"name": "retrieve_knowledge",
|
261 |
+
"description": "Retrieves knowledge from a database with a provided query.",
|
262 |
+
"parameters": {
|
263 |
+
"type": "object",
|
264 |
+
"properties": {
|
265 |
+
"query": {
|
266 |
+
"type": "string",
|
267 |
+
"description": "The query to search for in the vector store."
|
268 |
+
},
|
269 |
+
"n_results": {
|
270 |
+
"type": "integer",
|
271 |
+
"description": "The number of results to return. Default is 5."
|
272 |
+
},
|
273 |
+
"similarity_threshold": {
|
274 |
+
"type": "number",
|
275 |
+
"description": "The minimum similarity score for results. Default is 0.7."
|
276 |
+
}
|
277 |
+
},
|
278 |
+
"required": [
|
279 |
+
"query"
|
280 |
+
]
|
281 |
+
}
|
282 |
+
}
|
283 |
}
|
284 |
]
|