tdpeuter/open-webui
tdpeuter/open-webui
1
Fork 0
forked from open-webui/open-webui
Code Issues Pull requests 1 Projects 1 Releases 2 Packages 1 Activity Actions

Merge pull request #1687 from buroa/buroa/huggingface-embeddings

Browse source 
feat: move to native `sentence_transformers`
This commit is contained in:
Timothy Jaeryang Baek 2024-04-22 13:55:34 -07:00 • committed by GitHub
commit 0546ad58be
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 153 additions and 268 deletions
Download patch file Download diff file Expand all files Collapse all files

206
backend/apps/rag/main.py
View file

@ -13,7 +13,6 @@ import os, shutil, logging, re
from pathlib import Path
from typing import List
from chromadb.utils import embedding_functions
from chromadb.utils.batch_utils import create_batches
from langchain_community.document_loaders import (
@ -38,6 +37,7 @@ import mimetypes
import uuid
import json
import sentence_transformers
from apps.ollama.main import generate_ollama_embeddings, GenerateEmbeddingsForm
@ -48,11 +48,8 @@ from apps.web.models.documents import (
)
from apps.rag.utils import (
query_doc,
query_embeddings_doc,
query_collection,
query_embeddings_collection,
get_embedding_model_path,
generate_openai_embeddings,
)
@ -69,7 +66,7 @@ from config import (
DOCS_DIR,
RAG_EMBEDDING_ENGINE,
RAG_EMBEDDING_MODEL,
RAG_EMBEDDING_MODEL_AUTO_UPDATE,
RAG_EMBEDDING_MODEL_TRUST_REMOTE_CODE,
RAG_OPENAI_API_BASE_URL,
RAG_OPENAI_API_KEY,
DEVICE_TYPE,
@ -101,15 +98,12 @@ app.state.OPENAI_API_KEY = RAG_OPENAI_API_KEY
app.state.PDF_EXTRACT_IMAGES = False
app.state.sentence_transformer_ef = (
embedding_functions.SentenceTransformerEmbeddingFunction(
model_name=get_embedding_model_path(
app.state.RAG_EMBEDDING_MODEL, RAG_EMBEDDING_MODEL_AUTO_UPDATE
),
if app.state.RAG_EMBEDDING_ENGINE == "":
app.state.sentence_transformer_ef = sentence_transformers.SentenceTransformer(
app.state.RAG_EMBEDDING_MODEL,
device=DEVICE_TYPE,
trust_remote_code=RAG_EMBEDDING_MODEL_TRUST_REMOTE_CODE,
)
)
origins = ["*"]
@ -185,13 +179,10 @@ async def update_embedding_config(
app.state.OPENAI_API_BASE_URL = form_data.openai_config.url
app.state.OPENAI_API_KEY = form_data.openai_config.key
else:
sentence_transformer_ef = (
embedding_functions.SentenceTransformerEmbeddingFunction(
model_name=get_embedding_model_path(
form_data.embedding_model, True
),
device=DEVICE_TYPE,
)
sentence_transformer_ef = sentence_transformers.SentenceTransformer(
app.state.RAG_EMBEDDING_MODEL,
device=DEVICE_TYPE,
trust_remote_code=RAG_EMBEDDING_MODEL_TRUST_REMOTE_CODE,
)
app.state.RAG_EMBEDDING_MODEL = form_data.embedding_model
app.state.sentence_transformer_ef = sentence_transformer_ef
@ -294,38 +285,34 @@ def query_doc_handler(
form_data: QueryDocForm,
user=Depends(get_current_user),
):
try:
if app.state.RAG_EMBEDDING_ENGINE == "":
return query_doc(
collection_name=form_data.collection_name,
query=form_data.query,
k=form_data.k if form_data.k else app.state.TOP_K,
embedding_function=app.state.sentence_transformer_ef,
query_embeddings = app.state.sentence_transformer_ef.encode(
form_data.query
).tolist()
elif app.state.RAG_EMBEDDING_ENGINE == "ollama":
query_embeddings = generate_ollama_embeddings(
GenerateEmbeddingsForm(
**{
"model": app.state.RAG_EMBEDDING_MODEL,
"prompt": form_data.query,
}
)
)
elif app.state.RAG_EMBEDDING_ENGINE == "openai":
query_embeddings = generate_openai_embeddings(
model=app.state.RAG_EMBEDDING_MODEL,
text=form_data.query,
key=app.state.OPENAI_API_KEY,
url=app.state.OPENAI_API_BASE_URL,
)
else:
if app.state.RAG_EMBEDDING_ENGINE == "ollama":
query_embeddings = generate_ollama_embeddings(
GenerateEmbeddingsForm(
**{
"model": app.state.RAG_EMBEDDING_MODEL,
"prompt": form_data.query,
}
)
)
elif app.state.RAG_EMBEDDING_ENGINE == "openai":
query_embeddings = generate_openai_embeddings(
model=app.state.RAG_EMBEDDING_MODEL,
text=form_data.query,
key=app.state.OPENAI_API_KEY,
url=app.state.OPENAI_API_BASE_URL,
)
return query_embeddings_doc(
collection_name=form_data.collection_name,
query_embeddings=query_embeddings,
k=form_data.k if form_data.k else app.state.TOP_K,
)
return query_embeddings_doc(
collection_name=form_data.collection_name,
query=form_data.query,
query_embeddings=query_embeddings,
k=form_data.k if form_data.k else app.state.TOP_K,
)
except Exception as e:
log.exception(e)
@ -348,36 +335,31 @@ def query_collection_handler(
):
try:
if app.state.RAG_EMBEDDING_ENGINE == "":
return query_collection(
collection_names=form_data.collection_names,
query=form_data.query,
k=form_data.k if form_data.k else app.state.TOP_K,
embedding_function=app.state.sentence_transformer_ef,
)
else:
if app.state.RAG_EMBEDDING_ENGINE == "ollama":
query_embeddings = generate_ollama_embeddings(
GenerateEmbeddingsForm(
**{
"model": app.state.RAG_EMBEDDING_MODEL,
"prompt": form_data.query,
}
)
query_embeddings = app.state.sentence_transformer_ef.encode(
form_data.query
).tolist()
elif app.state.RAG_EMBEDDING_ENGINE == "ollama":
query_embeddings = generate_ollama_embeddings(
GenerateEmbeddingsForm(
**{
"model": app.state.RAG_EMBEDDING_MODEL,
"prompt": form_data.query,
}
)
elif app.state.RAG_EMBEDDING_ENGINE == "openai":
query_embeddings = generate_openai_embeddings(
model=app.state.RAG_EMBEDDING_MODEL,
text=form_data.query,
key=app.state.OPENAI_API_KEY,
url=app.state.OPENAI_API_BASE_URL,
)
return query_embeddings_collection(
collection_names=form_data.collection_names,
query_embeddings=query_embeddings,
k=form_data.k if form_data.k else app.state.TOP_K,
)
elif app.state.RAG_EMBEDDING_ENGINE == "openai":
query_embeddings = generate_openai_embeddings(
model=app.state.RAG_EMBEDDING_MODEL,
text=form_data.query,
key=app.state.OPENAI_API_KEY,
url=app.state.OPENAI_API_BASE_URL,
)
return query_embeddings_collection(
collection_names=form_data.collection_names,
query_embeddings=query_embeddings,
k=form_data.k if form_data.k else app.state.TOP_K,
)
except Exception as e:
log.exception(e)
@ -445,6 +427,8 @@ def store_docs_in_vector_db(docs, collection_name, overwrite: bool = False) -> b
log.info(f"store_docs_in_vector_db {docs} {collection_name}")
texts = [doc.page_content for doc in docs]
texts = list(map(lambda x: x.replace("\n", " "), texts))
metadatas = [doc.metadata for doc in docs]
try:
@ -454,52 +438,38 @@ def store_docs_in_vector_db(docs, collection_name, overwrite: bool = False) -> b
log.info(f"deleting existing collection {collection_name}")
CHROMA_CLIENT.delete_collection(name=collection_name)
collection = CHROMA_CLIENT.create_collection(name=collection_name)
if app.state.RAG_EMBEDDING_ENGINE == "":
collection = CHROMA_CLIENT.create_collection(
name=collection_name,
embedding_function=app.state.sentence_transformer_ef,
)
for batch in create_batches(
api=CHROMA_CLIENT,
ids=[str(uuid.uuid1()) for _ in texts],
metadatas=metadatas,
documents=texts,
):
collection.add(*batch)
else:
collection = CHROMA_CLIENT.create_collection(name=collection_name)
if app.state.RAG_EMBEDDING_ENGINE == "ollama":
embeddings = [
generate_ollama_embeddings(
GenerateEmbeddingsForm(
**{"model": app.state.RAG_EMBEDDING_MODEL, "prompt": text}
)
embeddings = app.state.sentence_transformer_ef.encode(texts).tolist()
elif app.state.RAG_EMBEDDING_ENGINE == "ollama":
embeddings = [
generate_ollama_embeddings(
GenerateEmbeddingsForm(
**{"model": app.state.RAG_EMBEDDING_MODEL, "prompt": text}
)
for text in texts
]
elif app.state.RAG_EMBEDDING_ENGINE == "openai":
embeddings = [
generate_openai_embeddings(
model=app.state.RAG_EMBEDDING_MODEL,
text=text,
key=app.state.OPENAI_API_KEY,
url=app.state.OPENAI_API_BASE_URL,
)
for text in texts
]
)
for text in texts
]
elif app.state.RAG_EMBEDDING_ENGINE == "openai":
embeddings = [
generate_openai_embeddings(
model=app.state.RAG_EMBEDDING_MODEL,
text=text,
key=app.state.OPENAI_API_KEY,
url=app.state.OPENAI_API_BASE_URL,
)
for text in texts
]
for batch in create_batches(
api=CHROMA_CLIENT,
ids=[str(uuid.uuid1()) for _ in texts],
metadatas=metadatas,
embeddings=embeddings,
documents=texts,
):
collection.add(*batch)
for batch in create_batches(
api=CHROMA_CLIENT,
ids=[str(uuid.uuid1()) for _ in texts],
metadatas=metadatas,
embeddings=embeddings,
documents=texts,
):
collection.add(*batch)
return True
except Exception as e:

182
backend/apps/rag/utils.py
View file

@ -1,13 +1,12 @@
import os
import re
import logging
from typing import List
import requests
from typing import List
from huggingface_hub import snapshot_download
from apps.ollama.main import generate_ollama_embeddings, GenerateEmbeddingsForm
from apps.ollama.main import (
generate_ollama_embeddings,
GenerateEmbeddingsForm,
)
from config import SRC_LOG_LEVELS, CHROMA_CLIENT
@ -16,29 +15,12 @@ log = logging.getLogger(__name__)
log.setLevel(SRC_LOG_LEVELS["RAG"])
def query_doc(collection_name: str, query: str, k: int, embedding_function):
try:
# if you use docker use the model from the environment variable
collection = CHROMA_CLIENT.get_collection(
name=collection_name,
embedding_function=embedding_function,
)
result = collection.query(
query_texts=[query],
n_results=k,
)
return result
except Exception as e:
raise e
def query_embeddings_doc(collection_name: str, query_embeddings, k: int):
def query_embeddings_doc(collection_name: str, query: str, query_embeddings, k: int):
try:
# if you use docker use the model from the environment variable
log.info(f"query_embeddings_doc {query_embeddings}")
collection = CHROMA_CLIENT.get_collection(
name=collection_name,
)
collection = CHROMA_CLIENT.get_collection(name=collection_name)
result = collection.query(
query_embeddings=[query_embeddings],
n_results=k,
@ -95,43 +77,20 @@ def merge_and_sort_query_results(query_results, k):
return merged_query_results
def query_collection(
collection_names: List[str], query: str, k: int, embedding_function
def query_embeddings_collection(
collection_names: List[str], query: str, query_embeddings, k: int
):
results = []
for collection_name in collection_names:
try:
# if you use docker use the model from the environment variable
collection = CHROMA_CLIENT.get_collection(
name=collection_name,
embedding_function=embedding_function,
)
result = collection.query(
query_texts=[query],
n_results=k,
)
results.append(result)
except:
pass
return merge_and_sort_query_results(results, k)
def query_embeddings_collection(collection_names: List[str], query_embeddings, k: int):
results = []
log.info(f"query_embeddings_collection {query_embeddings}")
for collection_name in collection_names:
try:
collection = CHROMA_CLIENT.get_collection(name=collection_name)
result = collection.query(
query_embeddings=[query_embeddings],
n_results=k,
result = query_embeddings_doc(
collection_name=collection_name,
query=query,
query_embeddings=query_embeddings,
k=k,
)
results.append(result)
except:
@ -197,51 +156,38 @@ def rag_messages(
context = doc["content"]
else:
if embedding_engine == "":
if doc["type"] == "collection":
context = query_collection(
collection_names=doc["collection_names"],
query=query,
k=k,
embedding_function=embedding_function,
)
else:
context = query_doc(
collection_name=doc["collection_name"],
query=query,
k=k,
embedding_function=embedding_function,
query_embeddings = embedding_function.encode(query).tolist()
elif embedding_engine == "ollama":
query_embeddings = generate_ollama_embeddings(
GenerateEmbeddingsForm(
**{
"model": embedding_model,
"prompt": query,
}
)
)
elif embedding_engine == "openai":
query_embeddings = generate_openai_embeddings(
model=embedding_model,
text=query,
key=openai_key,
url=openai_url,
)
if doc["type"] == "collection":
context = query_embeddings_collection(
collection_names=doc["collection_names"],
query=query,
query_embeddings=query_embeddings,
k=k,
)
else:
if embedding_engine == "ollama":
query_embeddings = generate_ollama_embeddings(
GenerateEmbeddingsForm(
**{
"model": embedding_model,
"prompt": query,
}
)
)
elif embedding_engine == "openai":
query_embeddings = generate_openai_embeddings(
model=embedding_model,
text=query,
key=openai_key,
url=openai_url,
)
if doc["type"] == "collection":
context = query_embeddings_collection(
collection_names=doc["collection_names"],
query_embeddings=query_embeddings,
k=k,
)
else:
context = query_embeddings_doc(
collection_name=doc["collection_name"],
query_embeddings=query_embeddings,
k=k,
)
context = query_embeddings_doc(
collection_name=doc["collection_name"],
query=query,
query_embeddings=query_embeddings,
k=k,
)
except Exception as e:
log.exception(e)
@ -283,46 +229,6 @@ def rag_messages(
return messages
def get_embedding_model_path(
embedding_model: str, update_embedding_model: bool = False
):
# Construct huggingface_hub kwargs with local_files_only to return the snapshot path
cache_dir = os.getenv("SENTENCE_TRANSFORMERS_HOME")
local_files_only = not update_embedding_model
snapshot_kwargs = {
"cache_dir": cache_dir,
"local_files_only": local_files_only,
}
log.debug(f"embedding_model: {embedding_model}")
log.debug(f"snapshot_kwargs: {snapshot_kwargs}")
# Inspiration from upstream sentence_transformers
if (
os.path.exists(embedding_model)
or ("\\" in embedding_model or embedding_model.count("/") > 1)
and local_files_only
):
# If fully qualified path exists, return input, else set repo_id
return embedding_model
elif "/" not in embedding_model:
# Set valid repo_id for model short-name
embedding_model = "sentence-transformers" + "/" + embedding_model
snapshot_kwargs["repo_id"] = embedding_model
# Attempt to query the huggingface_hub library to determine the local path and/or to update
try:
embedding_model_repo_path = snapshot_download(**snapshot_kwargs)
log.debug(f"embedding_model_repo_path: {embedding_model_repo_path}")
return embedding_model_repo_path
except Exception as e:
log.exception(f"Cannot determine embedding model snapshot path: {e}")
return embedding_model
def generate_openai_embeddings(
model: str, text: str, key: str, url: str = "https://api.openai.com/v1"
):