Spaces:
No application file
No application file
Create main.py
Browse files
main.py
ADDED
@@ -0,0 +1,208 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List, Tuple
|
2 |
+
from pathlib import Path
|
3 |
+
import os
|
4 |
+
import subprocess
|
5 |
+
|
6 |
+
from dotenv import load_dotenv
|
7 |
+
from haystack.preview import Pipeline
|
8 |
+
from haystack.preview.dataclasses import GeneratedAnswer
|
9 |
+
from haystack.preview.components.retrievers import MemoryBM25Retriever
|
10 |
+
from haystack.preview.components.generators.openai.gpt import GPTGenerator
|
11 |
+
from haystack.preview.components.builders.answer_builder import AnswerBuilder
|
12 |
+
from haystack.preview.components.builders.prompt_builder import PromptBuilder
|
13 |
+
from haystack.preview.components.preprocessors import (
|
14 |
+
DocumentCleaner,
|
15 |
+
TextDocumentSplitter,
|
16 |
+
)
|
17 |
+
from haystack.preview.components.writers import DocumentWriter
|
18 |
+
from haystack.preview.components.file_converters import TextFileToDocument
|
19 |
+
from haystack.preview.document_stores.memory import MemoryDocumentStore
|
20 |
+
import streamlit as st
|
21 |
+
|
22 |
+
# Load the environment variables, we're going to need it for OpenAI
|
23 |
+
load_dotenv()
|
24 |
+
|
25 |
+
# This is the list of documentation that we're going to fetch
|
26 |
+
DOCUMENTATIONS = [
|
27 |
+
(
|
28 |
+
"DocArray",
|
29 |
+
"https://github.com/docarray/docarray",
|
30 |
+
"./docs/**/*.md",
|
31 |
+
),
|
32 |
+
(
|
33 |
+
"Streamlit",
|
34 |
+
"https://github.com/streamlit/docs",
|
35 |
+
"./content/**/*.md",
|
36 |
+
),
|
37 |
+
(
|
38 |
+
"Jinja",
|
39 |
+
"https://github.com/pallets/jinja",
|
40 |
+
"./docs/**/*.rst",
|
41 |
+
),
|
42 |
+
(
|
43 |
+
"Pandas",
|
44 |
+
"https://github.com/pandas-dev/pandas",
|
45 |
+
"./doc/source/**/*.rst",
|
46 |
+
),
|
47 |
+
(
|
48 |
+
"Elasticsearch",
|
49 |
+
"https://github.com/elastic/elasticsearch",
|
50 |
+
"./docs/**/*.asciidoc",
|
51 |
+
),
|
52 |
+
(
|
53 |
+
"NumPy",
|
54 |
+
"https://github.com/numpy/numpy",
|
55 |
+
"./doc/**/*.rst",
|
56 |
+
),
|
57 |
+
]
|
58 |
+
|
59 |
+
DOCS_PATH = Path(__file__).parent / "downloaded_docs"
|
60 |
+
|
61 |
+
|
62 |
+
@st.cache_data(show_spinner=False)
|
63 |
+
def fetch(documentations: List[Tuple[str, str, str]]):
|
64 |
+
files = []
|
65 |
+
# Create the docs path if it doesn't exist
|
66 |
+
DOCS_PATH.mkdir(parents=True, exist_ok=True)
|
67 |
+
|
68 |
+
for name, url, pattern in documentations:
|
69 |
+
st.write(f"Fetching {name} repository")
|
70 |
+
repo = DOCS_PATH / name
|
71 |
+
# Attempt cloning only if it doesn't exist
|
72 |
+
if not repo.exists():
|
73 |
+
subprocess.run(["git", "clone", "--depth", "1", url, str(repo)], check=True)
|
74 |
+
res = subprocess.run(
|
75 |
+
["git", "rev-parse", "--abbrev-ref", "HEAD"],
|
76 |
+
check=True,
|
77 |
+
capture_output=True,
|
78 |
+
encoding="utf-8",
|
79 |
+
cwd=repo,
|
80 |
+
)
|
81 |
+
branch = res.stdout.strip()
|
82 |
+
for p in repo.glob(pattern):
|
83 |
+
data = {
|
84 |
+
"path": p,
|
85 |
+
"metadata": {
|
86 |
+
"url_source": f"{url}/tree/{branch}/{p.relative_to(repo)}",
|
87 |
+
"suffix": p.suffix,
|
88 |
+
},
|
89 |
+
}
|
90 |
+
files.append(data)
|
91 |
+
|
92 |
+
return files
|
93 |
+
|
94 |
+
|
95 |
+
@st.cache_resource(show_spinner=False)
|
96 |
+
def document_store():
|
97 |
+
# We're going to store the processed documents in here
|
98 |
+
return MemoryDocumentStore()
|
99 |
+
|
100 |
+
|
101 |
+
@st.cache_resource(show_spinner=False)
|
102 |
+
def index_files(files):
|
103 |
+
# We create some components
|
104 |
+
text_converter = TextFileToDocument(progress_bar=False)
|
105 |
+
document_cleaner = DocumentCleaner()
|
106 |
+
document_splitter = TextDocumentSplitter()
|
107 |
+
document_writer = DocumentWriter(
|
108 |
+
document_store=document_store(), policy="overwrite"
|
109 |
+
)
|
110 |
+
|
111 |
+
# And our pipeline
|
112 |
+
indexing_pipeline = Pipeline()
|
113 |
+
indexing_pipeline.add_component("converter", text_converter)
|
114 |
+
indexing_pipeline.add_component("cleaner", document_cleaner)
|
115 |
+
indexing_pipeline.add_component("splitter", document_splitter)
|
116 |
+
indexing_pipeline.add_component("writer", document_writer)
|
117 |
+
indexing_pipeline.connect("converter", "cleaner")
|
118 |
+
indexing_pipeline.connect("cleaner", "splitter")
|
119 |
+
indexing_pipeline.connect("splitter", "writer")
|
120 |
+
|
121 |
+
# And now we save the documentation in our MemoryDocumentStore
|
122 |
+
paths = []
|
123 |
+
metadata = []
|
124 |
+
for f in files:
|
125 |
+
paths.append(f["path"])
|
126 |
+
metadata.append(f["metadata"])
|
127 |
+
indexing_pipeline.run(
|
128 |
+
{
|
129 |
+
"converter": {
|
130 |
+
"paths": paths,
|
131 |
+
"metadata": metadata,
|
132 |
+
}
|
133 |
+
}
|
134 |
+
)
|
135 |
+
|
136 |
+
|
137 |
+
def search(question: str) -> GeneratedAnswer:
|
138 |
+
retriever = MemoryBM25Retriever(document_store=document_store(), top_k=5)
|
139 |
+
|
140 |
+
template = (
|
141 |
+
"Take a deep breath and think then answer given the context"
|
142 |
+
"Context: {{ documents|map(attribute='text')|replace('\n', ' ')|join(';') }}"
|
143 |
+
"Question: {{ query }}"
|
144 |
+
"Answer:"
|
145 |
+
)
|
146 |
+
prompt_builder = PromptBuilder(template)
|
147 |
+
|
148 |
+
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY", "")
|
149 |
+
generator = GPTGenerator(api_key=OPENAI_API_KEY)
|
150 |
+
answer_builder = AnswerBuilder()
|
151 |
+
|
152 |
+
query_pipeline = Pipeline()
|
153 |
+
|
154 |
+
query_pipeline.add_component("docs_retriever", retriever)
|
155 |
+
query_pipeline.add_component("prompt_builder", prompt_builder)
|
156 |
+
query_pipeline.add_component("gpt35", generator)
|
157 |
+
query_pipeline.add_component("answer_builder", answer_builder)
|
158 |
+
|
159 |
+
query_pipeline.connect("docs_retriever.documents", "prompt_builder.documents")
|
160 |
+
query_pipeline.connect("prompt_builder.prompt", "gpt35.prompt")
|
161 |
+
query_pipeline.connect("docs_retriever.documents", "answer_builder.documents")
|
162 |
+
query_pipeline.connect("gpt35.replies", "answer_builder.replies")
|
163 |
+
res = query_pipeline.run(
|
164 |
+
{
|
165 |
+
"docs_retriever": {"query": question},
|
166 |
+
"prompt_builder": {"query": question},
|
167 |
+
"answer_builder": {"query": question},
|
168 |
+
}
|
169 |
+
)
|
170 |
+
return res["answer_builder"]["answers"][0]
|
171 |
+
|
172 |
+
|
173 |
+
with st.status(
|
174 |
+
"Downloading documentation files...",
|
175 |
+
expanded=st.session_state.get("expanded", True),
|
176 |
+
) as status:
|
177 |
+
files = fetch(DOCUMENTATIONS)
|
178 |
+
status.update(label="Indexing documentation...")
|
179 |
+
index_files(files)
|
180 |
+
status.update(
|
181 |
+
label="Download and indexing complete!", state="complete", expanded=False
|
182 |
+
)
|
183 |
+
st.session_state["expanded"] = False
|
184 |
+
|
185 |
+
|
186 |
+
st.header("🔎 Documentation finder", divider="rainbow")
|
187 |
+
|
188 |
+
st.caption(
|
189 |
+
f"Use this to search answers for {', '.join([d[0] for d in DOCUMENTATIONS])}"
|
190 |
+
)
|
191 |
+
|
192 |
+
if question := st.text_input(
|
193 |
+
label="What do you need to know?", placeholder="What is a DataFrame?"
|
194 |
+
):
|
195 |
+
with st.spinner("Waiting"):
|
196 |
+
answer = search(question)
|
197 |
+
|
198 |
+
if not st.session_state.get("run_once", False):
|
199 |
+
st.balloons()
|
200 |
+
st.session_state["run_once"] = True
|
201 |
+
|
202 |
+
st.markdown(answer.data)
|
203 |
+
with st.expander("See sources:"):
|
204 |
+
for document in answer.documents:
|
205 |
+
url_source = document.metadata.get("url_source", "")
|
206 |
+
st.write(url_source)
|
207 |
+
st.text(document.text)
|
208 |
+
st.divider()
|