Spaces:
Runtime error
Runtime error
Merge pull request #10 from ugm2/feature/audio_output
Browse files- .gitignore +3 -1
- core/pipelines.py +34 -3
- core/search_index.py +11 -9
- interface/components.py +10 -3
- interface/pages.py +1 -1
- interface/utils.py +10 -0
- requirements.txt +5 -1
.gitignore
CHANGED
@@ -128,4 +128,6 @@ dmypy.json
|
|
128 |
# Pyre type checker
|
129 |
.pyre/
|
130 |
|
131 |
-
.vscode/
|
|
|
|
|
|
128 |
# Pyre type checker
|
129 |
.pyre/
|
130 |
|
131 |
+
.vscode/
|
132 |
+
|
133 |
+
data/audio/
|
core/pipelines.py
CHANGED
@@ -2,14 +2,20 @@
|
|
2 |
Haystack Pipelines
|
3 |
"""
|
4 |
|
|
|
5 |
from haystack import Pipeline
|
6 |
from haystack.document_stores import InMemoryDocumentStore
|
7 |
from haystack.nodes.retriever import DensePassageRetriever, TfidfRetriever
|
8 |
from haystack.nodes.preprocessor import PreProcessor
|
9 |
from haystack.nodes.ranker import SentenceTransformersRanker
|
|
|
|
|
10 |
|
|
|
|
|
11 |
|
12 |
-
|
|
|
13 |
"""
|
14 |
**Keyword Search Pipeline**
|
15 |
|
@@ -19,8 +25,6 @@ def keyword_search(index="documents", split_word_length=100):
|
|
19 |
|
20 |
- Documents that have more lexical overlap with the query are more likely to be relevant
|
21 |
- Words that occur in fewer documents are more significant than words that occur in many documents
|
22 |
-
|
23 |
-
:warning: **(HAYSTACK BUG) Keyword Search doesn't work if you reindex:** Please refresh page in order to reindex
|
24 |
"""
|
25 |
document_store = InMemoryDocumentStore(index=index)
|
26 |
keyword_retriever = TfidfRetriever(document_store=(document_store))
|
@@ -44,6 +48,15 @@ def keyword_search(index="documents", split_word_length=100):
|
|
44 |
document_store, name="DocumentStore", inputs=["Preprocessor"]
|
45 |
)
|
46 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
return search_pipeline, index_pipeline
|
48 |
|
49 |
|
@@ -52,6 +65,7 @@ def dense_passage_retrieval(
|
|
52 |
split_word_length=100,
|
53 |
query_embedding_model="facebook/dpr-question_encoder-single-nq-base",
|
54 |
passage_embedding_model="facebook/dpr-ctx_encoder-single-nq-base",
|
|
|
55 |
):
|
56 |
"""
|
57 |
**Dense Passage Retrieval Pipeline**
|
@@ -89,6 +103,15 @@ def dense_passage_retrieval(
|
|
89 |
document_store, name="DocumentStore", inputs=["DPRRetriever"]
|
90 |
)
|
91 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
92 |
return search_pipeline, index_pipeline
|
93 |
|
94 |
|
@@ -98,6 +121,7 @@ def dense_passage_retrieval_ranker(
|
|
98 |
query_embedding_model="facebook/dpr-question_encoder-single-nq-base",
|
99 |
passage_embedding_model="facebook/dpr-ctx_encoder-single-nq-base",
|
100 |
ranker_model="cross-encoder/ms-marco-MiniLM-L-12-v2",
|
|
|
101 |
):
|
102 |
"""
|
103 |
**Dense Passage Retrieval Ranker Pipeline**
|
@@ -118,4 +142,11 @@ def dense_passage_retrieval_ranker(
|
|
118 |
|
119 |
search_pipeline.add_node(ranker, name="Ranker", inputs=["DPRRetriever"])
|
120 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
121 |
return search_pipeline, index_pipeline
|
|
|
2 |
Haystack Pipelines
|
3 |
"""
|
4 |
|
5 |
+
from pathlib import Path
|
6 |
from haystack import Pipeline
|
7 |
from haystack.document_stores import InMemoryDocumentStore
|
8 |
from haystack.nodes.retriever import DensePassageRetriever, TfidfRetriever
|
9 |
from haystack.nodes.preprocessor import PreProcessor
|
10 |
from haystack.nodes.ranker import SentenceTransformersRanker
|
11 |
+
from haystack.nodes.audio.document_to_speech import DocumentToSpeech
|
12 |
+
import os
|
13 |
|
14 |
+
data_path = "data/"
|
15 |
+
os.makedirs(data_path, exist_ok=True)
|
16 |
|
17 |
+
|
18 |
+
def keyword_search(index="documents", split_word_length=100, audio_output=False):
|
19 |
"""
|
20 |
**Keyword Search Pipeline**
|
21 |
|
|
|
25 |
|
26 |
- Documents that have more lexical overlap with the query are more likely to be relevant
|
27 |
- Words that occur in fewer documents are more significant than words that occur in many documents
|
|
|
|
|
28 |
"""
|
29 |
document_store = InMemoryDocumentStore(index=index)
|
30 |
keyword_retriever = TfidfRetriever(document_store=(document_store))
|
|
|
48 |
document_store, name="DocumentStore", inputs=["Preprocessor"]
|
49 |
)
|
50 |
|
51 |
+
if audio_output:
|
52 |
+
doc2speech = DocumentToSpeech(
|
53 |
+
model_name_or_path="espnet/kan-bayashi_ljspeech_vits",
|
54 |
+
generated_audio_dir=Path(data_path + "audio"),
|
55 |
+
)
|
56 |
+
search_pipeline.add_node(
|
57 |
+
doc2speech, name="DocumentToSpeech", inputs=["TfidfRetriever"]
|
58 |
+
)
|
59 |
+
|
60 |
return search_pipeline, index_pipeline
|
61 |
|
62 |
|
|
|
65 |
split_word_length=100,
|
66 |
query_embedding_model="facebook/dpr-question_encoder-single-nq-base",
|
67 |
passage_embedding_model="facebook/dpr-ctx_encoder-single-nq-base",
|
68 |
+
audio_output=False,
|
69 |
):
|
70 |
"""
|
71 |
**Dense Passage Retrieval Pipeline**
|
|
|
103 |
document_store, name="DocumentStore", inputs=["DPRRetriever"]
|
104 |
)
|
105 |
|
106 |
+
if audio_output:
|
107 |
+
doc2speech = DocumentToSpeech(
|
108 |
+
model_name_or_path="espnet/kan-bayashi_ljspeech_vits",
|
109 |
+
generated_audio_dir=Path(data_path + "audio"),
|
110 |
+
)
|
111 |
+
search_pipeline.add_node(
|
112 |
+
doc2speech, name="DocumentToSpeech", inputs=["DPRRetriever"]
|
113 |
+
)
|
114 |
+
|
115 |
return search_pipeline, index_pipeline
|
116 |
|
117 |
|
|
|
121 |
query_embedding_model="facebook/dpr-question_encoder-single-nq-base",
|
122 |
passage_embedding_model="facebook/dpr-ctx_encoder-single-nq-base",
|
123 |
ranker_model="cross-encoder/ms-marco-MiniLM-L-12-v2",
|
124 |
+
audio_output=False,
|
125 |
):
|
126 |
"""
|
127 |
**Dense Passage Retrieval Ranker Pipeline**
|
|
|
142 |
|
143 |
search_pipeline.add_node(ranker, name="Ranker", inputs=["DPRRetriever"])
|
144 |
|
145 |
+
if audio_output:
|
146 |
+
doc2speech = DocumentToSpeech(
|
147 |
+
model_name_or_path="espnet/kan-bayashi_ljspeech_vits",
|
148 |
+
generated_audio_dir=Path(data_path + "audio"),
|
149 |
+
)
|
150 |
+
search_pipeline.add_node(doc2speech, name="DocumentToSpeech", inputs=["Ranker"])
|
151 |
+
|
152 |
return search_pipeline, index_pipeline
|
core/search_index.py
CHANGED
@@ -37,15 +37,17 @@ def search(queries, pipeline):
|
|
37 |
for res in matches:
|
38 |
if not score_is_empty:
|
39 |
score_is_empty = True if res.score is None else False
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
}
|
48 |
-
)
|
|
|
|
|
49 |
if not score_is_empty:
|
50 |
query_results = sorted(
|
51 |
query_results, key=lambda x: x["score"], reverse=True
|
|
|
37 |
for res in matches:
|
38 |
if not score_is_empty:
|
39 |
score_is_empty = True if res.score is None else False
|
40 |
+
match = {
|
41 |
+
"text": res.content,
|
42 |
+
"id": res.meta["id"],
|
43 |
+
"fragment_id": res.id,
|
44 |
+
"meta": res.meta,
|
45 |
+
}
|
46 |
+
if not score_is_empty:
|
47 |
+
match.update({"score": res.score})
|
48 |
+
if hasattr(res, "content_audio"):
|
49 |
+
match.update({"content_audio": res.content_audio})
|
50 |
+
query_results.append(match)
|
51 |
if not score_is_empty:
|
52 |
query_results = sorted(
|
53 |
query_results, key=lambda x: x["score"], reverse=True
|
interface/components.py
CHANGED
@@ -1,5 +1,10 @@
|
|
1 |
import streamlit as st
|
2 |
-
from interface.utils import
|
|
|
|
|
|
|
|
|
|
|
3 |
from interface.draw_pipelines import get_pipeline_graph
|
4 |
|
5 |
|
@@ -42,7 +47,7 @@ def component_select_pipeline(container):
|
|
42 |
"index_pipeline": index_pipeline,
|
43 |
"doc": pipeline_funcs[index_pipe].__doc__,
|
44 |
}
|
45 |
-
|
46 |
|
47 |
|
48 |
def component_show_pipeline(pipeline, pipeline_name):
|
@@ -65,8 +70,10 @@ def component_show_search_result(container, results):
|
|
65 |
st.markdown(f"**Document**: {document['id']}")
|
66 |
if "_split_id" in document["meta"]:
|
67 |
st.markdown(f"**Document Chunk**: {document['meta']['_split_id']}")
|
68 |
-
if
|
69 |
st.markdown(f"**Score**: {document['score']:.3f}")
|
|
|
|
|
70 |
st.markdown("---")
|
71 |
|
72 |
|
|
|
1 |
import streamlit as st
|
2 |
+
from interface.utils import (
|
3 |
+
get_pipelines,
|
4 |
+
extract_text_from_url,
|
5 |
+
extract_text_from_file,
|
6 |
+
reset_vars_data,
|
7 |
+
)
|
8 |
from interface.draw_pipelines import get_pipeline_graph
|
9 |
|
10 |
|
|
|
47 |
"index_pipeline": index_pipeline,
|
48 |
"doc": pipeline_funcs[index_pipe].__doc__,
|
49 |
}
|
50 |
+
reset_vars_data()
|
51 |
|
52 |
|
53 |
def component_show_pipeline(pipeline, pipeline_name):
|
|
|
70 |
st.markdown(f"**Document**: {document['id']}")
|
71 |
if "_split_id" in document["meta"]:
|
72 |
st.markdown(f"**Document Chunk**: {document['meta']['_split_id']}")
|
73 |
+
if "score" in document:
|
74 |
st.markdown(f"**Score**: {document['score']:.3f}")
|
75 |
+
if "content_audio" in document:
|
76 |
+
st.audio(str(document["content_audio"]))
|
77 |
st.markdown("---")
|
78 |
|
79 |
|
interface/pages.py
CHANGED
@@ -25,12 +25,12 @@ def page_landing_page(container):
|
|
25 |
"\n - Index raw text, URLs, CSVs, PDFs and Images"
|
26 |
"\n - Use Dense Passage Retrieval, Keyword Search pipeline and DPR Ranker pipelines"
|
27 |
"\n - Search the indexed documents"
|
|
|
28 |
)
|
29 |
st.markdown(
|
30 |
"TODO list:"
|
31 |
"\n - File type classification and converter nodes"
|
32 |
"\n - Audio to text support for indexing"
|
33 |
-
"\n - Include text to audio to read responses"
|
34 |
"\n - Build other pipelines"
|
35 |
)
|
36 |
st.markdown(
|
|
|
25 |
"\n - Index raw text, URLs, CSVs, PDFs and Images"
|
26 |
"\n - Use Dense Passage Retrieval, Keyword Search pipeline and DPR Ranker pipelines"
|
27 |
"\n - Search the indexed documents"
|
28 |
+
"\n - Read your responses out loud using the `audio_output` option!"
|
29 |
)
|
30 |
st.markdown(
|
31 |
"TODO list:"
|
32 |
"\n - File type classification and converter nodes"
|
33 |
"\n - Audio to text support for indexing"
|
|
|
34 |
"\n - Build other pipelines"
|
35 |
)
|
36 |
st.markdown(
|
interface/utils.py
CHANGED
@@ -1,5 +1,8 @@
|
|
1 |
from io import StringIO
|
|
|
|
|
2 |
import core.pipelines as pipelines_functions
|
|
|
3 |
from inspect import getmembers, isfunction, signature
|
4 |
from newspaper import Article
|
5 |
from PyPDF2 import PdfFileReader
|
@@ -23,6 +26,13 @@ def get_pipelines():
|
|
23 |
return pipeline_names, pipeline_funcs, pipeline_func_parameters
|
24 |
|
25 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
@st.experimental_memo
|
27 |
def extract_text_from_url(url: str):
|
28 |
article = Article(url)
|
|
|
1 |
from io import StringIO
|
2 |
+
import os
|
3 |
+
import shutil
|
4 |
import core.pipelines as pipelines_functions
|
5 |
+
from core.pipelines import data_path
|
6 |
from inspect import getmembers, isfunction, signature
|
7 |
from newspaper import Article
|
8 |
from PyPDF2 import PdfFileReader
|
|
|
26 |
return pipeline_names, pipeline_funcs, pipeline_func_parameters
|
27 |
|
28 |
|
29 |
+
def reset_vars_data():
|
30 |
+
st.session_state["doc_id"] = 0
|
31 |
+
# Delete data files
|
32 |
+
shutil.rmtree(data_path)
|
33 |
+
os.makedirs(data_path, exist_ok=True)
|
34 |
+
|
35 |
+
|
36 |
@st.experimental_memo
|
37 |
def extract_text_from_url(url: str):
|
38 |
article = Article(url)
|
requirements.txt
CHANGED
@@ -5,4 +5,8 @@ black==22.8.0
|
|
5 |
plotly==5.10.0
|
6 |
newspaper3k==0.2.8
|
7 |
PyPDF2==2.10.7
|
8 |
-
pytesseract==0.3.10
|
|
|
|
|
|
|
|
|
|
5 |
plotly==5.10.0
|
6 |
newspaper3k==0.2.8
|
7 |
PyPDF2==2.10.7
|
8 |
+
pytesseract==0.3.10
|
9 |
+
soundfile==0.10.3.post1
|
10 |
+
espnet
|
11 |
+
pydub==0.25.1
|
12 |
+
espnet_model_zoo==0.1.7
|