Spaces:
Running
Running
return embeddings from storage retrieval
Browse files- document_qa/document_qa_engine.py +295 -77
- requirements.txt +11 -11
- streamlit_app.py +24 -20
document_qa/document_qa_engine.py
CHANGED
|
@@ -1,23 +1,43 @@
|
|
| 1 |
import copy
|
| 2 |
import os
|
| 3 |
from pathlib import Path
|
| 4 |
-
from typing import Union, Any
|
| 5 |
|
| 6 |
import tiktoken
|
| 7 |
-
from grobid_client.grobid_client import GrobidClient
|
| 8 |
from langchain.chains import create_extraction_chain
|
| 9 |
from langchain.chains.question_answering import load_qa_chain, stuff_prompt, refine_prompts, map_reduce_prompt, \
|
| 10 |
map_rerank_prompt
|
| 11 |
from langchain.prompts import SystemMessagePromptTemplate, HumanMessagePromptTemplate, ChatPromptTemplate
|
| 12 |
from langchain.retrievers import MultiQueryRetriever
|
| 13 |
from langchain.schema import Document
|
| 14 |
-
from
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
from tqdm import tqdm
|
| 16 |
|
| 17 |
from document_qa.grobid_processors import GrobidProcessor
|
| 18 |
|
| 19 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
class TextMerger:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
def __init__(self, model_name=None, encoding_name="gpt2"):
|
| 22 |
if model_name is not None:
|
| 23 |
self.enc = tiktoken.encoding_for_model(model_name)
|
|
@@ -85,52 +105,187 @@ class TextMerger:
|
|
| 85 |
|
| 86 |
return new_passages_struct
|
| 87 |
|
| 88 |
-
class DataStorage:
|
| 89 |
|
|
|
|
| 90 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 91 |
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 96 |
embeddings_dict = {}
|
| 97 |
embeddings_map_from_md5 = {}
|
| 98 |
embeddings_map_to_md5 = {}
|
| 99 |
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
llm,
|
| 109 |
-
embedding_function,
|
| 110 |
-
qa_chain_type="stuff",
|
| 111 |
-
embeddings_root_path=None,
|
| 112 |
-
grobid_url=None,
|
| 113 |
-
memory=None
|
| 114 |
-
):
|
| 115 |
self.embedding_function = embedding_function
|
| 116 |
-
self.llm = llm
|
| 117 |
-
self.memory = memory
|
| 118 |
-
self.chain = load_qa_chain(llm, chain_type=qa_chain_type)
|
| 119 |
-
self.text_merger = TextMerger()
|
| 120 |
|
| 121 |
-
if
|
| 122 |
-
self.embeddings_root_path =
|
| 123 |
-
if not os.path.exists(
|
| 124 |
-
os.makedirs(
|
| 125 |
else:
|
| 126 |
self.load_embeddings(self.embeddings_root_path)
|
| 127 |
|
| 128 |
-
if grobid_url:
|
| 129 |
-
self.grobid_processor = GrobidProcessor(grobid_url)
|
| 130 |
-
|
| 131 |
def load_embeddings(self, embeddings_root_path: Union[str, Path]) -> None:
|
| 132 |
"""
|
| 133 |
-
Load the
|
| 134 |
The root path of the embeddings containing one data store for each document in each subdirectory
|
| 135 |
"""
|
| 136 |
|
|
@@ -141,8 +296,10 @@ class DocumentQAEngine:
|
|
| 141 |
return
|
| 142 |
|
| 143 |
for embedding_document_dir in embeddings_directories:
|
| 144 |
-
self.embeddings_dict[embedding_document_dir.name] =
|
| 145 |
-
|
|
|
|
|
|
|
| 146 |
|
| 147 |
filename_list = list(Path(embedding_document_dir).glob('*.storage_filename'))
|
| 148 |
if filename_list:
|
|
@@ -161,9 +318,60 @@ class DocumentQAEngine:
|
|
| 161 |
def get_filename_from_md5(self, md5):
|
| 162 |
return self.embeddings_map_from_md5[md5]
|
| 163 |
|
| 164 |
-
def
|
| 165 |
-
|
| 166 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 167 |
# self.load_embeddings(self.embeddings_root_path)
|
| 168 |
|
| 169 |
if verbose:
|
|
@@ -192,16 +400,22 @@ class DocumentQAEngine:
|
|
| 192 |
else:
|
| 193 |
return None, response, coordinates
|
| 194 |
|
| 195 |
-
def query_storage(self, query: str, doc_id, context_size=4):
|
| 196 |
-
|
|
|
|
|
|
|
|
|
|
| 197 |
|
| 198 |
context_as_text = [doc.page_content for doc in documents]
|
| 199 |
-
return context_as_text
|
| 200 |
|
| 201 |
def query_storage_and_embeddings(self, query: str, doc_id, context_size=4):
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
|
|
|
|
|
|
|
|
|
| 205 |
|
| 206 |
context_as_text = [doc.page_content for doc in relevant_documents]
|
| 207 |
return context_as_text
|
|
@@ -229,11 +443,11 @@ class DocumentQAEngine:
|
|
| 229 |
|
| 230 |
return parsed_output
|
| 231 |
|
| 232 |
-
def _run_query(self, doc_id, query, context_size=4):
|
| 233 |
relevant_documents = self._get_context(doc_id, query, context_size)
|
| 234 |
relevant_document_coordinates = [doc.metadata['coordinates'].split(";") if 'coordinates' in doc.metadata else []
|
| 235 |
for doc in
|
| 236 |
-
relevant_documents]
|
| 237 |
response = self.chain.run(input_documents=relevant_documents,
|
| 238 |
question=query)
|
| 239 |
|
|
@@ -241,33 +455,40 @@ class DocumentQAEngine:
|
|
| 241 |
self.memory.save_context({"input": query}, {"output": response})
|
| 242 |
return response, relevant_document_coordinates
|
| 243 |
|
| 244 |
-
def _get_context(self, doc_id, query, context_size=4):
|
| 245 |
-
db = self.embeddings_dict[doc_id]
|
| 246 |
retriever = db.as_retriever(search_kwargs={"k": context_size})
|
| 247 |
relevant_documents = retriever.get_relevant_documents(query)
|
|
|
|
|
|
|
|
|
|
| 248 |
if self.memory and len(self.memory.buffer_as_messages) > 0:
|
| 249 |
relevant_documents.append(
|
| 250 |
Document(
|
| 251 |
page_content="""Following, the previous question and answers. Use these information only when in the question there are unspecified references:\n{}\n\n""".format(
|
| 252 |
self.memory.buffer_as_str))
|
| 253 |
)
|
| 254 |
-
return relevant_documents
|
| 255 |
|
| 256 |
-
def
|
| 257 |
-
"""
|
| 258 |
-
|
|
|
|
|
|
|
| 259 |
docs = db.get()
|
| 260 |
return docs['documents']
|
| 261 |
|
| 262 |
def _get_context_multiquery(self, doc_id, query, context_size=4):
|
| 263 |
-
db = self.embeddings_dict[doc_id].as_retriever(search_kwargs={"k": context_size})
|
| 264 |
multi_query_retriever = MultiQueryRetriever.from_llm(retriever=db, llm=self.llm)
|
| 265 |
relevant_documents = multi_query_retriever.get_relevant_documents(query)
|
| 266 |
return relevant_documents
|
| 267 |
|
| 268 |
def get_text_from_document(self, pdf_file_path, chunk_size=-1, perc_overlap=0.1, verbose=False):
|
| 269 |
"""
|
| 270 |
-
Extract text from documents using Grobid
|
|
|
|
|
|
|
| 271 |
"""
|
| 272 |
if verbose:
|
| 273 |
print("File", pdf_file_path)
|
|
@@ -307,7 +528,13 @@ class DocumentQAEngine:
|
|
| 307 |
|
| 308 |
return texts, metadatas, ids
|
| 309 |
|
| 310 |
-
def create_memory_embeddings(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 311 |
texts, metadata, ids = self.get_text_from_document(
|
| 312 |
pdf_path,
|
| 313 |
chunk_size=chunk_size,
|
|
@@ -317,25 +544,17 @@ class DocumentQAEngine:
|
|
| 317 |
else:
|
| 318 |
hash = metadata[0]['hash']
|
| 319 |
|
| 320 |
-
|
| 321 |
-
self.embeddings_dict[hash] = Chroma.from_texts(texts,
|
| 322 |
-
embedding=self.embedding_function,
|
| 323 |
-
metadatas=metadata,
|
| 324 |
-
collection_name=hash)
|
| 325 |
-
else:
|
| 326 |
-
# if 'documents' in self.embeddings_dict[hash].get() and len(self.embeddings_dict[hash].get()['documents']) == 0:
|
| 327 |
-
# self.embeddings_dict[hash].delete(ids=self.embeddings_dict[hash].get()['ids'])
|
| 328 |
-
self.embeddings_dict[hash].delete_collection()
|
| 329 |
-
self.embeddings_dict[hash] = Chroma.from_texts(texts,
|
| 330 |
-
embedding=self.embedding_function,
|
| 331 |
-
metadatas=metadata,
|
| 332 |
-
collection_name=hash)
|
| 333 |
-
|
| 334 |
-
self.embeddings_root_path = None
|
| 335 |
|
| 336 |
return hash
|
| 337 |
|
| 338 |
-
def create_embeddings(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 339 |
input_files = []
|
| 340 |
for root, dirs, files in os.walk(pdfs_dir_path, followlinks=False):
|
| 341 |
for file_ in files:
|
|
@@ -347,17 +566,16 @@ class DocumentQAEngine:
|
|
| 347 |
desc="Grobid + embeddings processing"):
|
| 348 |
|
| 349 |
md5 = self.calculate_md5(input_file)
|
| 350 |
-
data_path = os.path.join(self.embeddings_root_path, md5)
|
| 351 |
|
| 352 |
if os.path.exists(data_path):
|
| 353 |
print(data_path, "exists. Skipping it ")
|
| 354 |
continue
|
| 355 |
-
include = ["biblio"] if include_biblio else []
|
| 356 |
texts, metadata, ids = self.get_text_from_document(
|
| 357 |
input_file,
|
| 358 |
chunk_size=chunk_size,
|
| 359 |
-
perc_overlap=perc_overlap
|
| 360 |
-
include=include)
|
| 361 |
filename = metadata[0]['filename']
|
| 362 |
|
| 363 |
vector_db_document = Chroma.from_texts(texts,
|
|
|
|
| 1 |
import copy
|
| 2 |
import os
|
| 3 |
from pathlib import Path
|
| 4 |
+
from typing import Union, Any, Optional, List, Dict, Tuple, ClassVar, Collection
|
| 5 |
|
| 6 |
import tiktoken
|
|
|
|
| 7 |
from langchain.chains import create_extraction_chain
|
| 8 |
from langchain.chains.question_answering import load_qa_chain, stuff_prompt, refine_prompts, map_reduce_prompt, \
|
| 9 |
map_rerank_prompt
|
| 10 |
from langchain.prompts import SystemMessagePromptTemplate, HumanMessagePromptTemplate, ChatPromptTemplate
|
| 11 |
from langchain.retrievers import MultiQueryRetriever
|
| 12 |
from langchain.schema import Document
|
| 13 |
+
from langchain_community.vectorstores.chroma import Chroma, DEFAULT_K
|
| 14 |
+
from langchain_community.vectorstores.faiss import FAISS
|
| 15 |
+
from langchain_core.callbacks import CallbackManagerForRetrieverRun
|
| 16 |
+
from langchain_core.utils import xor_args
|
| 17 |
+
from langchain_core.vectorstores import VectorStore, VectorStoreRetriever
|
| 18 |
from tqdm import tqdm
|
| 19 |
|
| 20 |
from document_qa.grobid_processors import GrobidProcessor
|
| 21 |
|
| 22 |
|
| 23 |
+
def _results_to_docs_scores_and_embeddings(results: Any) -> List[Tuple[Document, float, List[float]]]:
|
| 24 |
+
return [
|
| 25 |
+
(Document(page_content=result[0], metadata=result[1] or {}), result[2], result[3])
|
| 26 |
+
for result in zip(
|
| 27 |
+
results["documents"][0],
|
| 28 |
+
results["metadatas"][0],
|
| 29 |
+
results["distances"][0],
|
| 30 |
+
results["embeddings"][0],
|
| 31 |
+
)
|
| 32 |
+
]
|
| 33 |
+
|
| 34 |
+
|
| 35 |
class TextMerger:
|
| 36 |
+
"""
|
| 37 |
+
This class tries to replicate the RecursiveTextSplitter from LangChain, to preserve and merge the
|
| 38 |
+
coordinate information from the PDF document.
|
| 39 |
+
"""
|
| 40 |
+
|
| 41 |
def __init__(self, model_name=None, encoding_name="gpt2"):
|
| 42 |
if model_name is not None:
|
| 43 |
self.enc = tiktoken.encoding_for_model(model_name)
|
|
|
|
| 105 |
|
| 106 |
return new_passages_struct
|
| 107 |
|
|
|
|
| 108 |
|
| 109 |
+
class BaseRetrieval:
|
| 110 |
|
| 111 |
+
def __init__(
|
| 112 |
+
self,
|
| 113 |
+
persist_directory: Path,
|
| 114 |
+
embedding_function
|
| 115 |
+
):
|
| 116 |
+
self.embedding_function = embedding_function
|
| 117 |
+
self.persist_directory = persist_directory
|
| 118 |
+
|
| 119 |
+
|
| 120 |
+
class AdvancedVectorStoreRetriever(VectorStoreRetriever):
|
| 121 |
+
allowed_search_types: ClassVar[Collection[str]] = (
|
| 122 |
+
"similarity",
|
| 123 |
+
"similarity_score_threshold",
|
| 124 |
+
"mmr",
|
| 125 |
+
"similarity_with_embeddings"
|
| 126 |
+
)
|
| 127 |
+
|
| 128 |
+
def _get_relevant_documents(
|
| 129 |
+
self, query: str, *, run_manager: CallbackManagerForRetrieverRun
|
| 130 |
+
) -> List[Document]:
|
| 131 |
+
if self.search_type == "similarity":
|
| 132 |
+
docs = self.vectorstore.similarity_search(query, **self.search_kwargs)
|
| 133 |
+
elif self.search_type == "similarity_score_threshold":
|
| 134 |
+
docs_and_similarities = (
|
| 135 |
+
self.vectorstore.similarity_search_with_relevance_scores(
|
| 136 |
+
query, **self.search_kwargs
|
| 137 |
+
)
|
| 138 |
+
)
|
| 139 |
+
for doc, similarity in docs_and_similarities:
|
| 140 |
+
if '__similarity' not in doc.metadata.keys():
|
| 141 |
+
doc.metadata['__similarity'] = similarity
|
| 142 |
+
|
| 143 |
+
docs = [doc for doc, _ in docs_and_similarities]
|
| 144 |
+
elif self.search_type == "mmr":
|
| 145 |
+
docs = self.vectorstore.max_marginal_relevance_search(
|
| 146 |
+
query, **self.search_kwargs
|
| 147 |
+
)
|
| 148 |
+
elif self.search_type == "similarity_with_embeddings":
|
| 149 |
+
docs_scores_and_embeddings = (
|
| 150 |
+
self.vectorstore.advanced_similarity_search(
|
| 151 |
+
query, **self.search_kwargs
|
| 152 |
+
)
|
| 153 |
+
)
|
| 154 |
|
| 155 |
+
for doc, score, embeddings in docs_scores_and_embeddings:
|
| 156 |
+
if '__embeddings' not in doc.metadata.keys():
|
| 157 |
+
doc.metadata['__embeddings'] = embeddings
|
| 158 |
+
if '__similarity' not in doc.metadata.keys():
|
| 159 |
+
doc.metadata['__similarity'] = score
|
| 160 |
+
|
| 161 |
+
docs = [doc for doc, _, _ in docs_scores_and_embeddings]
|
| 162 |
+
else:
|
| 163 |
+
raise ValueError(f"search_type of {self.search_type} not allowed.")
|
| 164 |
+
return docs
|
| 165 |
+
|
| 166 |
+
|
| 167 |
+
class AdvancedVectorStore(VectorStore):
|
| 168 |
+
def as_retriever(self, **kwargs: Any) -> AdvancedVectorStoreRetriever:
|
| 169 |
+
tags = kwargs.pop("tags", None) or []
|
| 170 |
+
tags.extend(self._get_retriever_tags())
|
| 171 |
+
return AdvancedVectorStoreRetriever(vectorstore=self, **kwargs, tags=tags)
|
| 172 |
+
|
| 173 |
+
|
| 174 |
+
class ChromaAdvancedRetrieval(Chroma, AdvancedVectorStore):
|
| 175 |
+
def __init__(self, **kwargs):
|
| 176 |
+
super().__init__(**kwargs)
|
| 177 |
+
|
| 178 |
+
@xor_args(("query_texts", "query_embeddings"))
|
| 179 |
+
def __query_collection(
|
| 180 |
+
self,
|
| 181 |
+
query_texts: Optional[List[str]] = None,
|
| 182 |
+
query_embeddings: Optional[List[List[float]]] = None,
|
| 183 |
+
n_results: int = 4,
|
| 184 |
+
where: Optional[Dict[str, str]] = None,
|
| 185 |
+
where_document: Optional[Dict[str, str]] = None,
|
| 186 |
+
**kwargs: Any,
|
| 187 |
+
) -> List[Document]:
|
| 188 |
+
"""Query the chroma collection."""
|
| 189 |
+
try:
|
| 190 |
+
import chromadb # noqa: F401
|
| 191 |
+
except ImportError:
|
| 192 |
+
raise ValueError(
|
| 193 |
+
"Could not import chromadb python package. "
|
| 194 |
+
"Please install it with `pip install chromadb`."
|
| 195 |
+
)
|
| 196 |
+
return self._collection.query(
|
| 197 |
+
query_texts=query_texts,
|
| 198 |
+
query_embeddings=query_embeddings,
|
| 199 |
+
n_results=n_results,
|
| 200 |
+
where=where,
|
| 201 |
+
where_document=where_document,
|
| 202 |
+
**kwargs,
|
| 203 |
+
)
|
| 204 |
+
|
| 205 |
+
def advanced_similarity_search(
|
| 206 |
+
self,
|
| 207 |
+
query: str,
|
| 208 |
+
k: int = DEFAULT_K,
|
| 209 |
+
filter: Optional[Dict[str, str]] = None,
|
| 210 |
+
**kwargs: Any,
|
| 211 |
+
) -> [List[Document], float, List[float]]:
|
| 212 |
+
docs_scores_and_embeddings = self.similarity_search_with_scores_and_embeddings(query, k, filter=filter)
|
| 213 |
+
return docs_scores_and_embeddings
|
| 214 |
+
|
| 215 |
+
def similarity_search_with_scores_and_embeddings(
|
| 216 |
+
self,
|
| 217 |
+
query: str,
|
| 218 |
+
k: int = DEFAULT_K,
|
| 219 |
+
filter: Optional[Dict[str, str]] = None,
|
| 220 |
+
where_document: Optional[Dict[str, str]] = None,
|
| 221 |
+
**kwargs: Any,
|
| 222 |
+
) -> List[Tuple[Document, float, List[float]]]:
|
| 223 |
+
|
| 224 |
+
if self._embedding_function is None:
|
| 225 |
+
results = self.__query_collection(
|
| 226 |
+
query_texts=[query],
|
| 227 |
+
n_results=k,
|
| 228 |
+
where=filter,
|
| 229 |
+
where_document=where_document,
|
| 230 |
+
include=['metadatas', 'documents', 'embeddings', 'distances']
|
| 231 |
+
)
|
| 232 |
+
else:
|
| 233 |
+
query_embedding = self._embedding_function.embed_query(query)
|
| 234 |
+
results = self.__query_collection(
|
| 235 |
+
query_embeddings=[query_embedding],
|
| 236 |
+
n_results=k,
|
| 237 |
+
where=filter,
|
| 238 |
+
where_document=where_document,
|
| 239 |
+
include=['metadatas', 'documents', 'embeddings', 'distances']
|
| 240 |
+
)
|
| 241 |
+
|
| 242 |
+
return _results_to_docs_scores_and_embeddings(results)
|
| 243 |
+
|
| 244 |
+
|
| 245 |
+
class FAISSAdvancedRetrieval(FAISS):
|
| 246 |
+
pass
|
| 247 |
+
|
| 248 |
+
|
| 249 |
+
class NER_Retrival(VectorStore):
|
| 250 |
+
"""
|
| 251 |
+
This class implement a retrieval based on NER models.
|
| 252 |
+
This is an alternative retrieval to embeddings that relies on extracted entities.
|
| 253 |
+
"""
|
| 254 |
+
pass
|
| 255 |
+
|
| 256 |
+
|
| 257 |
+
engines = {
|
| 258 |
+
'chroma': ChromaAdvancedRetrieval,
|
| 259 |
+
'faiss': FAISSAdvancedRetrieval,
|
| 260 |
+
'ner': NER_Retrival
|
| 261 |
+
}
|
| 262 |
+
|
| 263 |
+
|
| 264 |
+
class DataStorage:
|
| 265 |
embeddings_dict = {}
|
| 266 |
embeddings_map_from_md5 = {}
|
| 267 |
embeddings_map_to_md5 = {}
|
| 268 |
|
| 269 |
+
def __init__(
|
| 270 |
+
self,
|
| 271 |
+
embedding_function,
|
| 272 |
+
root_path: Path = None,
|
| 273 |
+
engine=ChromaAdvancedRetrieval,
|
| 274 |
+
) -> None:
|
| 275 |
+
self.root_path = root_path
|
| 276 |
+
self.engine = engine
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 277 |
self.embedding_function = embedding_function
|
|
|
|
|
|
|
|
|
|
|
|
|
| 278 |
|
| 279 |
+
if root_path is not None:
|
| 280 |
+
self.embeddings_root_path = root_path
|
| 281 |
+
if not os.path.exists(root_path):
|
| 282 |
+
os.makedirs(root_path)
|
| 283 |
else:
|
| 284 |
self.load_embeddings(self.embeddings_root_path)
|
| 285 |
|
|
|
|
|
|
|
|
|
|
| 286 |
def load_embeddings(self, embeddings_root_path: Union[str, Path]) -> None:
|
| 287 |
"""
|
| 288 |
+
Load the vector storage assuming they are all persisted and stored in a single directory.
|
| 289 |
The root path of the embeddings containing one data store for each document in each subdirectory
|
| 290 |
"""
|
| 291 |
|
|
|
|
| 296 |
return
|
| 297 |
|
| 298 |
for embedding_document_dir in embeddings_directories:
|
| 299 |
+
self.embeddings_dict[embedding_document_dir.name] = self.engine(
|
| 300 |
+
persist_directory=embedding_document_dir.path,
|
| 301 |
+
embedding_function=self.embedding_function
|
| 302 |
+
)
|
| 303 |
|
| 304 |
filename_list = list(Path(embedding_document_dir).glob('*.storage_filename'))
|
| 305 |
if filename_list:
|
|
|
|
| 318 |
def get_filename_from_md5(self, md5):
|
| 319 |
return self.embeddings_map_from_md5[md5]
|
| 320 |
|
| 321 |
+
def embed_document(self, doc_id, texts, metadatas):
|
| 322 |
+
if doc_id not in self.embeddings_dict.keys():
|
| 323 |
+
self.embeddings_dict[doc_id] = self.engine.from_texts(texts,
|
| 324 |
+
embedding=self.embedding_function,
|
| 325 |
+
metadatas=metadatas,
|
| 326 |
+
collection_name=doc_id)
|
| 327 |
+
else:
|
| 328 |
+
# Workaround Chroma (?) breaking change
|
| 329 |
+
self.embeddings_dict[doc_id].delete_collection()
|
| 330 |
+
self.embeddings_dict[doc_id] = self.engine.from_texts(texts,
|
| 331 |
+
embedding=self.embedding_function,
|
| 332 |
+
metadatas=metadatas,
|
| 333 |
+
collection_name=doc_id)
|
| 334 |
+
|
| 335 |
+
self.embeddings_root_path = None
|
| 336 |
+
|
| 337 |
+
|
| 338 |
+
class DocumentQAEngine:
|
| 339 |
+
llm = None
|
| 340 |
+
qa_chain_type = None
|
| 341 |
+
|
| 342 |
+
default_prompts = {
|
| 343 |
+
'stuff': stuff_prompt,
|
| 344 |
+
'refine': refine_prompts,
|
| 345 |
+
"map_reduce": map_reduce_prompt,
|
| 346 |
+
"map_rerank": map_rerank_prompt
|
| 347 |
+
}
|
| 348 |
+
|
| 349 |
+
def __init__(self,
|
| 350 |
+
llm,
|
| 351 |
+
data_storage: DataStorage,
|
| 352 |
+
qa_chain_type="stuff",
|
| 353 |
+
grobid_url=None,
|
| 354 |
+
memory=None
|
| 355 |
+
):
|
| 356 |
+
|
| 357 |
+
self.llm = llm
|
| 358 |
+
self.memory = memory
|
| 359 |
+
self.chain = load_qa_chain(llm, chain_type=qa_chain_type)
|
| 360 |
+
self.text_merger = TextMerger()
|
| 361 |
+
self.data_storage = data_storage
|
| 362 |
+
|
| 363 |
+
if grobid_url:
|
| 364 |
+
self.grobid_processor = GrobidProcessor(grobid_url)
|
| 365 |
+
|
| 366 |
+
def query_document(
|
| 367 |
+
self,
|
| 368 |
+
query: str,
|
| 369 |
+
doc_id,
|
| 370 |
+
output_parser=None,
|
| 371 |
+
context_size=4,
|
| 372 |
+
extraction_schema=None,
|
| 373 |
+
verbose=False
|
| 374 |
+
) -> (Any, str):
|
| 375 |
# self.load_embeddings(self.embeddings_root_path)
|
| 376 |
|
| 377 |
if verbose:
|
|
|
|
| 400 |
else:
|
| 401 |
return None, response, coordinates
|
| 402 |
|
| 403 |
+
def query_storage(self, query: str, doc_id, context_size=4) -> (List[Document], list):
|
| 404 |
+
"""
|
| 405 |
+
Returns the context related to a given query
|
| 406 |
+
"""
|
| 407 |
+
documents, coordinates = self._get_context(doc_id, query, context_size)
|
| 408 |
|
| 409 |
context_as_text = [doc.page_content for doc in documents]
|
| 410 |
+
return context_as_text, coordinates
|
| 411 |
|
| 412 |
def query_storage_and_embeddings(self, query: str, doc_id, context_size=4):
|
| 413 |
+
"""
|
| 414 |
+
Returns both the context and the embedding information from a given query
|
| 415 |
+
"""
|
| 416 |
+
db = self.data_storage.embeddings_dict[doc_id]
|
| 417 |
+
retriever = db.as_retriever(search_kwargs={"k": context_size}, search_type="similarity_with_embeddings")
|
| 418 |
+
relevant_documents = retriever.get_relevant_documents(query)
|
| 419 |
|
| 420 |
context_as_text = [doc.page_content for doc in relevant_documents]
|
| 421 |
return context_as_text
|
|
|
|
| 443 |
|
| 444 |
return parsed_output
|
| 445 |
|
| 446 |
+
def _run_query(self, doc_id, query, context_size=4) -> (List[Document], list):
|
| 447 |
relevant_documents = self._get_context(doc_id, query, context_size)
|
| 448 |
relevant_document_coordinates = [doc.metadata['coordinates'].split(";") if 'coordinates' in doc.metadata else []
|
| 449 |
for doc in
|
| 450 |
+
relevant_documents]
|
| 451 |
response = self.chain.run(input_documents=relevant_documents,
|
| 452 |
question=query)
|
| 453 |
|
|
|
|
| 455 |
self.memory.save_context({"input": query}, {"output": response})
|
| 456 |
return response, relevant_document_coordinates
|
| 457 |
|
| 458 |
+
def _get_context(self, doc_id, query, context_size=4) -> (List[Document], list):
|
| 459 |
+
db = self.data_storage.embeddings_dict[doc_id]
|
| 460 |
retriever = db.as_retriever(search_kwargs={"k": context_size})
|
| 461 |
relevant_documents = retriever.get_relevant_documents(query)
|
| 462 |
+
relevant_document_coordinates = [doc.metadata['coordinates'].split(";") if 'coordinates' in doc.metadata else []
|
| 463 |
+
for doc in
|
| 464 |
+
relevant_documents]
|
| 465 |
if self.memory and len(self.memory.buffer_as_messages) > 0:
|
| 466 |
relevant_documents.append(
|
| 467 |
Document(
|
| 468 |
page_content="""Following, the previous question and answers. Use these information only when in the question there are unspecified references:\n{}\n\n""".format(
|
| 469 |
self.memory.buffer_as_str))
|
| 470 |
)
|
| 471 |
+
return relevant_documents, relevant_document_coordinates
|
| 472 |
|
| 473 |
+
def get_full_context_by_document(self, doc_id):
|
| 474 |
+
"""
|
| 475 |
+
Return the full context from the document
|
| 476 |
+
"""
|
| 477 |
+
db = self.data_storage.embeddings_dict[doc_id]
|
| 478 |
docs = db.get()
|
| 479 |
return docs['documents']
|
| 480 |
|
| 481 |
def _get_context_multiquery(self, doc_id, query, context_size=4):
|
| 482 |
+
db = self.data_storage.embeddings_dict[doc_id].as_retriever(search_kwargs={"k": context_size})
|
| 483 |
multi_query_retriever = MultiQueryRetriever.from_llm(retriever=db, llm=self.llm)
|
| 484 |
relevant_documents = multi_query_retriever.get_relevant_documents(query)
|
| 485 |
return relevant_documents
|
| 486 |
|
| 487 |
def get_text_from_document(self, pdf_file_path, chunk_size=-1, perc_overlap=0.1, verbose=False):
|
| 488 |
"""
|
| 489 |
+
Extract text from documents using Grobid.
|
| 490 |
+
- if chunk_size is < 0, keeps each paragraph separately
|
| 491 |
+
- if chunk_size > 0, aggregate all paragraphs and split them again using an approximate chunk size
|
| 492 |
"""
|
| 493 |
if verbose:
|
| 494 |
print("File", pdf_file_path)
|
|
|
|
| 528 |
|
| 529 |
return texts, metadatas, ids
|
| 530 |
|
| 531 |
+
def create_memory_embeddings(
|
| 532 |
+
self,
|
| 533 |
+
pdf_path,
|
| 534 |
+
doc_id=None,
|
| 535 |
+
chunk_size=500,
|
| 536 |
+
perc_overlap=0.1
|
| 537 |
+
):
|
| 538 |
texts, metadata, ids = self.get_text_from_document(
|
| 539 |
pdf_path,
|
| 540 |
chunk_size=chunk_size,
|
|
|
|
| 544 |
else:
|
| 545 |
hash = metadata[0]['hash']
|
| 546 |
|
| 547 |
+
self.data_storage.embed_document(hash, texts, metadata)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 548 |
|
| 549 |
return hash
|
| 550 |
|
| 551 |
+
def create_embeddings(
|
| 552 |
+
self,
|
| 553 |
+
pdfs_dir_path: Path,
|
| 554 |
+
chunk_size=500,
|
| 555 |
+
perc_overlap=0.1,
|
| 556 |
+
include_biblio=False
|
| 557 |
+
):
|
| 558 |
input_files = []
|
| 559 |
for root, dirs, files in os.walk(pdfs_dir_path, followlinks=False):
|
| 560 |
for file_ in files:
|
|
|
|
| 566 |
desc="Grobid + embeddings processing"):
|
| 567 |
|
| 568 |
md5 = self.calculate_md5(input_file)
|
| 569 |
+
data_path = os.path.join(self.data_storage.embeddings_root_path, md5)
|
| 570 |
|
| 571 |
if os.path.exists(data_path):
|
| 572 |
print(data_path, "exists. Skipping it ")
|
| 573 |
continue
|
| 574 |
+
# include = ["biblio"] if include_biblio else []
|
| 575 |
texts, metadata, ids = self.get_text_from_document(
|
| 576 |
input_file,
|
| 577 |
chunk_size=chunk_size,
|
| 578 |
+
perc_overlap=perc_overlap)
|
|
|
|
| 579 |
filename = metadata[0]['filename']
|
| 580 |
|
| 581 |
vector_db_document = Chroma.from_texts(texts,
|
requirements.txt
CHANGED
|
@@ -4,10 +4,10 @@ grobid-client-python==0.0.7
|
|
| 4 |
grobid_tei_xml==0.1.3
|
| 5 |
|
| 6 |
# Utils
|
| 7 |
-
tqdm==4.66.
|
| 8 |
pyyaml==6.0.1
|
| 9 |
-
pytest==
|
| 10 |
-
streamlit==1.
|
| 11 |
lxml
|
| 12 |
Beautifulsoup4
|
| 13 |
python-dotenv
|
|
@@ -15,13 +15,13 @@ watchdog
|
|
| 15 |
dateparser
|
| 16 |
|
| 17 |
# LLM
|
| 18 |
-
chromadb==0.4.
|
| 19 |
-
tiktoken==0.
|
| 20 |
-
openai==
|
| 21 |
-
langchain==0.
|
| 22 |
-
langchain-core==0.1.
|
| 23 |
typing-inspect==0.9.0
|
| 24 |
-
typing_extensions==4.
|
| 25 |
-
pydantic==2.4
|
| 26 |
-
sentence_transformers==2.
|
| 27 |
streamlit-pdf-viewer
|
|
|
|
| 4 |
grobid_tei_xml==0.1.3
|
| 5 |
|
| 6 |
# Utils
|
| 7 |
+
tqdm==4.66.2
|
| 8 |
pyyaml==6.0.1
|
| 9 |
+
pytest==8.1.1
|
| 10 |
+
streamlit==1.33.0
|
| 11 |
lxml
|
| 12 |
Beautifulsoup4
|
| 13 |
python-dotenv
|
|
|
|
| 15 |
dateparser
|
| 16 |
|
| 17 |
# LLM
|
| 18 |
+
chromadb==0.4.24
|
| 19 |
+
tiktoken==0.6.0
|
| 20 |
+
openai==1.16.2
|
| 21 |
+
langchain==0.1.14
|
| 22 |
+
langchain-core==0.1.40
|
| 23 |
typing-inspect==0.9.0
|
| 24 |
+
typing_extensions==4.11.0
|
| 25 |
+
pydantic==2.6.4
|
| 26 |
+
sentence_transformers==2.6.1
|
| 27 |
streamlit-pdf-viewer
|
streamlit_app.py
CHANGED
|
@@ -9,15 +9,16 @@ from langchain.llms.huggingface_hub import HuggingFaceHub
|
|
| 9 |
from langchain.memory import ConversationBufferWindowMemory
|
| 10 |
from streamlit_pdf_viewer import pdf_viewer
|
| 11 |
|
|
|
|
|
|
|
| 12 |
dotenv.load_dotenv(override=True)
|
| 13 |
|
| 14 |
import streamlit as st
|
| 15 |
from langchain.chat_models import ChatOpenAI
|
| 16 |
from langchain.embeddings import OpenAIEmbeddings, HuggingFaceEmbeddings
|
| 17 |
|
| 18 |
-
from document_qa.document_qa_engine import DocumentQAEngine
|
| 19 |
from document_qa.grobid_processors import GrobidAggregationProcessor, decorate_text_with_annotations
|
| 20 |
-
from grobid_client_generic import GrobidClientGeneric
|
| 21 |
|
| 22 |
OPENAI_MODELS = ['gpt-3.5-turbo',
|
| 23 |
"gpt-4",
|
|
@@ -168,14 +169,15 @@ def init_qa(model, api_key=None):
|
|
| 168 |
st.stop()
|
| 169 |
return
|
| 170 |
|
| 171 |
-
|
|
|
|
| 172 |
|
| 173 |
|
| 174 |
@st.cache_resource
|
| 175 |
def init_ner():
|
| 176 |
quantities_client = QuantitiesAPI(os.environ['GROBID_QUANTITIES_URL'], check_server=True)
|
| 177 |
|
| 178 |
-
materials_client =
|
| 179 |
config_materials = {
|
| 180 |
'grobid': {
|
| 181 |
"server": os.environ['GROBID_MATERIALS_URL'],
|
|
@@ -190,10 +192,8 @@ def init_ner():
|
|
| 190 |
|
| 191 |
materials_client.set_config(config_materials)
|
| 192 |
|
| 193 |
-
gqa = GrobidAggregationProcessor(
|
| 194 |
-
|
| 195 |
-
grobid_superconductors_client=materials_client
|
| 196 |
-
)
|
| 197 |
return gqa
|
| 198 |
|
| 199 |
|
|
@@ -340,9 +340,12 @@ with st.sidebar:
|
|
| 340 |
|
| 341 |
st.session_state['pdf_rendering'] = st.radio(
|
| 342 |
"PDF rendering mode",
|
| 343 |
-
|
| 344 |
-
index=
|
| 345 |
disabled=not uploaded_file,
|
|
|
|
|
|
|
|
|
|
| 346 |
)
|
| 347 |
|
| 348 |
st.divider()
|
|
@@ -441,7 +444,8 @@ with right_column:
|
|
| 441 |
text_response = None
|
| 442 |
if mode == "Embeddings":
|
| 443 |
with st.spinner("Generating LLM response..."):
|
| 444 |
-
text_response = st.session_state['rqa'][model].
|
|
|
|
| 445 |
context_size=context_size)
|
| 446 |
elif mode == "LLM":
|
| 447 |
with st.spinner("Generating response..."):
|
|
@@ -449,14 +453,14 @@ with right_column:
|
|
| 449 |
st.session_state.doc_id,
|
| 450 |
context_size=context_size)
|
| 451 |
|
| 452 |
-
|
| 453 |
-
|
| 454 |
-
|
| 455 |
-
|
| 456 |
-
|
| 457 |
-
|
| 458 |
-
|
| 459 |
-
|
| 460 |
|
| 461 |
if not text_response:
|
| 462 |
st.error("Something went wrong. Contact Luca Foppiano (Foppiano.Luca@nims.co.jp) to report the issue.")
|
|
@@ -486,5 +490,5 @@ with left_column:
|
|
| 486 |
height=800,
|
| 487 |
annotation_outline_size=1,
|
| 488 |
annotations=st.session_state['annotations'],
|
| 489 |
-
rendering=
|
| 490 |
)
|
|
|
|
| 9 |
from langchain.memory import ConversationBufferWindowMemory
|
| 10 |
from streamlit_pdf_viewer import pdf_viewer
|
| 11 |
|
| 12 |
+
from document_qa.ner_client_generic import NERClientGeneric
|
| 13 |
+
|
| 14 |
dotenv.load_dotenv(override=True)
|
| 15 |
|
| 16 |
import streamlit as st
|
| 17 |
from langchain.chat_models import ChatOpenAI
|
| 18 |
from langchain.embeddings import OpenAIEmbeddings, HuggingFaceEmbeddings
|
| 19 |
|
| 20 |
+
from document_qa.document_qa_engine import DocumentQAEngine, DataStorage
|
| 21 |
from document_qa.grobid_processors import GrobidAggregationProcessor, decorate_text_with_annotations
|
|
|
|
| 22 |
|
| 23 |
OPENAI_MODELS = ['gpt-3.5-turbo',
|
| 24 |
"gpt-4",
|
|
|
|
| 169 |
st.stop()
|
| 170 |
return
|
| 171 |
|
| 172 |
+
storage = DataStorage(embeddings)
|
| 173 |
+
return DocumentQAEngine(chat, storage, grobid_url=os.environ['GROBID_URL'], memory=st.session_state['memory'])
|
| 174 |
|
| 175 |
|
| 176 |
@st.cache_resource
|
| 177 |
def init_ner():
|
| 178 |
quantities_client = QuantitiesAPI(os.environ['GROBID_QUANTITIES_URL'], check_server=True)
|
| 179 |
|
| 180 |
+
materials_client = NERClientGeneric(ping=True)
|
| 181 |
config_materials = {
|
| 182 |
'grobid': {
|
| 183 |
"server": os.environ['GROBID_MATERIALS_URL'],
|
|
|
|
| 192 |
|
| 193 |
materials_client.set_config(config_materials)
|
| 194 |
|
| 195 |
+
gqa = GrobidAggregationProcessor(grobid_quantities_client=quantities_client,
|
| 196 |
+
grobid_superconductors_client=materials_client)
|
|
|
|
|
|
|
| 197 |
return gqa
|
| 198 |
|
| 199 |
|
|
|
|
| 340 |
|
| 341 |
st.session_state['pdf_rendering'] = st.radio(
|
| 342 |
"PDF rendering mode",
|
| 343 |
+
("unwrap", "legacy_embed"),
|
| 344 |
+
index=0,
|
| 345 |
disabled=not uploaded_file,
|
| 346 |
+
help="PDF rendering engine."
|
| 347 |
+
"Note: The Legacy PDF viewer does not support annotations and might not work on Chrome.",
|
| 348 |
+
format_func=lambda q: "Legacy PDF Viewer" if q == "legacy_embed" else "Streamlit PDF Viewer (Pdf.js)"
|
| 349 |
)
|
| 350 |
|
| 351 |
st.divider()
|
|
|
|
| 444 |
text_response = None
|
| 445 |
if mode == "Embeddings":
|
| 446 |
with st.spinner("Generating LLM response..."):
|
| 447 |
+
text_response, coordinates = st.session_state['rqa'][model].query_storage(question,
|
| 448 |
+
st.session_state.doc_id,
|
| 449 |
context_size=context_size)
|
| 450 |
elif mode == "LLM":
|
| 451 |
with st.spinner("Generating response..."):
|
|
|
|
| 453 |
st.session_state.doc_id,
|
| 454 |
context_size=context_size)
|
| 455 |
|
| 456 |
+
annotations = [[GrobidAggregationProcessor.box_to_dict([cs for cs in c.split(",")]) for c in coord_doc]
|
| 457 |
+
for coord_doc in coordinates]
|
| 458 |
+
gradients = generate_color_gradient(len(annotations))
|
| 459 |
+
for i, color in enumerate(gradients):
|
| 460 |
+
for annotation in annotations[i]:
|
| 461 |
+
annotation['color'] = color
|
| 462 |
+
st.session_state['annotations'] = [annotation for annotation_doc in annotations for annotation in
|
| 463 |
+
annotation_doc]
|
| 464 |
|
| 465 |
if not text_response:
|
| 466 |
st.error("Something went wrong. Contact Luca Foppiano (Foppiano.Luca@nims.co.jp) to report the issue.")
|
|
|
|
| 490 |
height=800,
|
| 491 |
annotation_outline_size=1,
|
| 492 |
annotations=st.session_state['annotations'],
|
| 493 |
+
rendering=st.session_state['pdf_rendering']
|
| 494 |
)
|