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

feat: hybrid search and reranking support

Browse source
This commit is contained in:
Steven Kreitzer 2024-04-22 18:36:46 -05:00
parent db801aee79
commit c0259aad67
10 changed files with 262 additions and 131 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
CHANGELOG.md
View file

@ -10,6 +10,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
### Added
- **🛠️ Improved Embedding Model Support**: You can now use any embedding model `sentence_transformers` supports.
- **🌟 Enhanced RAG Pipeline**: Added `BM25` hybrid searching with reranking model support using `sentence_transformers`.
## [0.1.120] - 2024-04-20

7
Dockerfile
View file

@ -10,7 +10,7 @@ ARG USE_CUDA_VER=cu121
# for better performance and multilangauge support use "intfloat/multilingual-e5-large" (~2.5GB) or "intfloat/multilingual-e5-base" (~1.5GB)
# IMPORTANT: If you change the embedding model (sentence-transformers/all-MiniLM-L6-v2) and vice versa, you aren't able to use RAG Chat with your previous documents loaded in the WebUI! You need to re-embed them.
ARG USE_EMBEDDING_MODEL=sentence-transformers/all-MiniLM-L6-v2
ARG USE_RERANKING_MODEL=BAAI/bge-reranker-base
ARG USE_RERANKING_MODEL=""
######## WebUI frontend ########
FROM --platform=$BUILDPLATFORM node:21-alpine3.19 as build
@ -67,6 +67,9 @@ ENV WHISPER_MODEL="base" \
ENV RAG_EMBEDDING_MODEL="$USE_EMBEDDING_MODEL_DOCKER" \
RAG_RERANKING_MODEL="$USE_RERANKING_MODEL_DOCKER" \
SENTENCE_TRANSFORMERS_HOME="/app/backend/data/cache/embedding/models"
## Hugging Face download cache ##
ENV HF_HOME="/app/backend/data/cache/embedding/models"
#### Other models ##########################################################
WORKDIR /app/backend
@ -102,13 +105,11 @@ RUN pip3 install uv && \
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/$USE_CUDA_DOCKER_VER --no-cache-dir && \
uv pip install --system -r requirements.txt --no-cache-dir && \
python -c "import os; from sentence_transformers import SentenceTransformer; SentenceTransformer(os.environ['RAG_EMBEDDING_MODEL'], device='cpu')" && \
python -c "import os; from sentence_transformers import CrossEncoder; CrossEncoder(os.environ['RAG_RERANKING_MODEL'], device='cpu')" && \
python -c "import os; from faster_whisper import WhisperModel; WhisperModel(os.environ['WHISPER_MODEL'], device='cpu', compute_type='int8', download_root=os.environ['WHISPER_MODEL_DIR'])"; \
else \
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu --no-cache-dir && \
uv pip install --system -r requirements.txt --no-cache-dir && \
python -c "import os; from sentence_transformers import SentenceTransformer; SentenceTransformer(os.environ['RAG_EMBEDDING_MODEL'], device='cpu')" && \
python -c "import os; from sentence_transformers import CrossEncoder; CrossEncoder(os.environ['RAG_RERANKING_MODEL'], device='cpu')" && \
python -c "import os; from faster_whisper import WhisperModel; WhisperModel(os.environ['WHISPER_MODEL'], device='cpu', compute_type='int8', download_root=os.environ['WHISPER_MODEL_DIR'])"; \
fi

4
backend/apps/ollama/main.py
View file

@ -92,6 +92,10 @@ async def get_ollama_api_urls(user=Depends(get_admin_user)):
return {"OLLAMA_BASE_URLS": app.state.OLLAMA_BASE_URLS}
def get_ollama_endpoint(url_idx: int = 0):
return app.state.OLLAMA_BASE_URLS[url_idx]
class UrlUpdateForm(BaseModel):
urls: List[str]

42
backend/apps/rag/main.py
View file

@ -64,6 +64,8 @@ from config import (
SRC_LOG_LEVELS,
UPLOAD_DIR,
DOCS_DIR,
RAG_TOP_K,
RAG_RELEVANCE_THRESHOLD,
RAG_EMBEDDING_ENGINE,
RAG_EMBEDDING_MODEL,
RAG_EMBEDDING_MODEL_TRUST_REMOTE_CODE,
@ -86,7 +88,8 @@ log.setLevel(SRC_LOG_LEVELS["RAG"])
app = FastAPI()
app.state.TOP_K = 4
app.state.TOP_K = RAG_TOP_K
app.state.RELEVANCE_THRESHOLD = RAG_RELEVANCE_THRESHOLD
app.state.CHUNK_SIZE = CHUNK_SIZE
app.state.CHUNK_OVERLAP = CHUNK_OVERLAP
@ -107,12 +110,17 @@ if app.state.RAG_EMBEDDING_ENGINE == "":
device=DEVICE_TYPE,
trust_remote_code=RAG_EMBEDDING_MODEL_TRUST_REMOTE_CODE,
)
else:
app.state.sentence_transformer_ef = None
app.state.sentence_transformer_rf = sentence_transformers.CrossEncoder(
if not app.state.RAG_RERANKING_MODEL == "":
app.state.sentence_transformer_rf = sentence_transformers.CrossEncoder(
app.state.RAG_RERANKING_MODEL,
device=DEVICE_TYPE,
trust_remote_code=RAG_RERANKING_MODEL_TRUST_REMOTE_CODE,
)
)
else:
app.state.sentence_transformer_rf = None
origins = ["*"]
@ -185,22 +193,22 @@ async def update_embedding_config(
)
try:
app.state.RAG_EMBEDDING_ENGINE = form_data.embedding_engine
app.state.RAG_EMBEDDING_MODEL = form_data.embedding_model
if app.state.RAG_EMBEDDING_ENGINE in ["ollama", "openai"]:
app.state.RAG_EMBEDDING_MODEL = form_data.embedding_model
app.state.sentence_transformer_ef = None
if form_data.openai_config != None:
app.state.OPENAI_API_BASE_URL = form_data.openai_config.url
app.state.OPENAI_API_KEY = form_data.openai_config.key
app.state.sentence_transformer_ef = None
else:
sentence_transformer_ef = sentence_transformers.SentenceTransformer(
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,
)
app.state.RAG_EMBEDDING_MODEL = form_data.embedding_model
app.state.sentence_transformer_ef = sentence_transformer_ef
)
return {
"status": True,
@ -233,6 +241,10 @@ async def update_reranking_config(
)
try:
app.state.RAG_RERANKING_MODEL = form_data.reranking_model
if app.state.RAG_RERANKING_MODEL == "":
app.state.sentence_transformer_rf = None
else:
app.state.sentence_transformer_rf = sentence_transformers.CrossEncoder(
app.state.RAG_RERANKING_MODEL,
device=DEVICE_TYPE,
@ -302,11 +314,13 @@ async def get_query_settings(user=Depends(get_admin_user)):
"status": True,
"template": app.state.RAG_TEMPLATE,
"k": app.state.TOP_K,
"r": app.state.RELEVANCE_THRESHOLD,
}
class QuerySettingsForm(BaseModel):
k: Optional[int] = None
r: Optional[float] = None
template: Optional[str] = None
@ -316,6 +330,7 @@ async def update_query_settings(
):
app.state.RAG_TEMPLATE = form_data.template if form_data.template else RAG_TEMPLATE
app.state.TOP_K = form_data.k if form_data.k else 4
app.state.RELEVANCE_THRESHOLD = form_data.r if form_data.r else 0.0
return {"status": True, "template": app.state.RAG_TEMPLATE}
@ -323,6 +338,7 @@ class QueryDocForm(BaseModel):
collection_name: str
query: str
k: Optional[int] = None
r: Optional[float] = None
@app.post("/query/doc")
@ -343,6 +359,7 @@ def query_doc_handler(
collection_name=form_data.collection_name,
query=form_data.query,
k=form_data.k if form_data.k else app.state.TOP_K,
r=form_data.r if form_data.r else app.state.RELEVANCE_THRESHOLD,
embeddings_function=embeddings_function,
reranking_function=app.state.sentence_transformer_rf,
)
@ -358,6 +375,7 @@ class QueryCollectionsForm(BaseModel):
collection_names: List[str]
query: str
k: Optional[int] = None
r: Optional[float] = None
@app.post("/query/collection")
@ -378,6 +396,7 @@ def query_collection_handler(
collection_names=form_data.collection_names,
query=form_data.query,
k=form_data.k if form_data.k else app.state.TOP_K,
r=form_data.r if form_data.r else app.state.RELEVANCE_THRESHOLD,
embeddings_function=embeddings_function,
reranking_function=app.state.sentence_transformer_rf,
)
@ -467,12 +486,7 @@ def store_docs_in_vector_db(docs, collection_name, overwrite: bool = False) -> b
)
embedding_texts = list(map(lambda x: x.replace("\n", " "), texts))
if app.state.RAG_EMBEDDING_ENGINE == "":
embeddings = embedding_func(embedding_texts)
else:
embeddings = [
embedding_func(embedding_texts) for text in texts
]
for batch in create_batches(
api=CHROMA_CLIENT,

189
backend/apps/rag/utils.py
View file

@ -1,8 +1,5 @@
import logging
import requests
import operator
import sentence_transformers
from typing import List
@ -11,8 +8,10 @@ from apps.ollama.main import (
GenerateEmbeddingsForm,
)
from langchain_core.documents import Document
from langchain_community.retrievers import BM25Retriever
from langchain.retrievers import (
BM25Retriever,
ContextualCompressionRetriever,
EnsembleRetriever,
)
@ -27,6 +26,7 @@ def query_embeddings_doc(
collection_name: str,
query: str,
k: int,
r: float,
embeddings_function,
reranking_function,
):
@ -34,7 +34,6 @@ def query_embeddings_doc(
# if you use docker use the model from the environment variable
collection = CHROMA_CLIENT.get_collection(name=collection_name)
# keyword search
documents = collection.get() # get all documents
bm25_retriever = BM25Retriever.from_texts(
texts=documents.get("documents"),
@ -42,30 +41,32 @@ def query_embeddings_doc(
)
bm25_retriever.k = k
# semantic search (vector)
chroma_retriever = ChromaRetriever(
collection=collection,
k=k,
embeddings_function=embeddings_function,
top_n=k,
)
# hybrid search (ensemble)
ensemble_retriever = EnsembleRetriever(
retrievers=[bm25_retriever, chroma_retriever],
weights=[0.6, 0.4]
retrievers=[bm25_retriever, chroma_retriever], weights=[0.5, 0.5]
)
documents = ensemble_retriever.invoke(query)
result = query_results_rank(
query=query,
documents=documents,
k=k,
compressor = RerankCompressor(
embeddings_function=embeddings_function,
reranking_function=reranking_function,
r_score=r,
top_n=k,
)
compression_retriever = ContextualCompressionRetriever(
base_compressor=compressor, base_retriever=ensemble_retriever
)
result = compression_retriever.invoke(query)
result = {
"distances": [[d[1].item() for d in result]],
"documents": [[d[0].page_content for d in result]],
"metadatas": [[d[0].metadata for d in result]],
"distances": [[d.metadata.get("score") for d in result]],
"documents": [[d.page_content for d in result]],
"metadatas": [[d.metadata for d in result]],
}
return result
@ -73,33 +74,29 @@ def query_embeddings_doc(
raise e
def query_results_rank(query: str, documents, k: int, reranking_function):
scores = reranking_function.predict([(query, doc.page_content) for doc in documents])
docs_with_scores = list(zip(documents, scores))
result = sorted(docs_with_scores, key=operator.itemgetter(1), reverse=True)
return result[: k]
def merge_and_sort_query_results(query_results, k):
# Initialize lists to store combined data
combined_distances = []
combined_documents = []
combined_metadatas = []
# Combine data from each dictionary
for data in query_results:
combined_distances.extend(data["distances"][0])
combined_documents.extend(data["documents"][0])
combined_metadatas.extend(data["metadatas"][0])
# Create a list of tuples (distance, document, metadata)
combined = list(
zip(combined_distances, combined_documents, combined_metadatas)
)
combined = list(zip(combined_distances, combined_documents, combined_metadatas))
# Sort the list based on distances
combined.sort(key=lambda x: x[0])
# We don't have anything :-(
if not combined:
sorted_distances = []
sorted_documents = []
sorted_metadatas = []
else:
# Unzip the sorted list
sorted_distances, sorted_documents, sorted_metadatas = zip(*combined)
@ -109,22 +106,20 @@ def merge_and_sort_query_results(query_results, k):
sorted_metadatas = list(sorted_metadatas)[:k]
# Create the output dictionary
merged_query_results = {
result = {
"distances": [sorted_distances],
"documents": [sorted_documents],
"metadatas": [sorted_metadatas],
"embeddings": None,
"uris": None,
"data": None,
}
return merged_query_results
return result
def query_embeddings_collection(
collection_names: List[str],
query: str,
k: int,
r: float,
embeddings_function,
reranking_function,
):
@ -137,6 +132,7 @@ def query_embeddings_collection(
collection_name=collection_name,
query=query,
k=k,
r=r,
embeddings_function=embeddings_function,
reranking_function=reranking_function,
)
@ -162,8 +158,9 @@ def query_embeddings_function(
):
if embedding_engine == "":
return lambda query: embedding_function.encode(query).tolist()
elif embedding_engine == "ollama":
return lambda query: generate_ollama_embeddings(
elif embedding_engine in ["ollama", "openai"]:
if embedding_engine == "ollama":
func = lambda query: generate_ollama_embeddings(
GenerateEmbeddingsForm(
**{
"model": embedding_model,
@ -172,19 +169,28 @@ def query_embeddings_function(
)
)
elif embedding_engine == "openai":
return lambda query: generate_openai_embeddings(
func = lambda query: generate_openai_embeddings(
model=embedding_model,
text=query,
key=openai_key,
url=openai_url,
)
def generate_multiple(query, f):
if isinstance(query, list):
return [f(q) for q in query]
else:
return f(query)
return lambda query: generate_multiple(query, func)
def rag_messages(
docs,
messages,
template,
k,
r,
embedding_engine,
embedding_model,
embedding_function,
@ -221,16 +227,6 @@ def rag_messages(
content_type = None
query = ""
relevant_contexts = []
for doc in docs:
context = None
try:
if doc["type"] == "text":
context = doc["content"]
else:
embeddings_function = query_embeddings_function(
embedding_engine,
embedding_model,
@ -239,11 +235,32 @@ def rag_messages(
openai_url,
)
if doc["type"] == "collection":
extracted_collections = []
relevant_contexts = []
for doc in docs:
context = None
collection = doc.get("collection_name")
if collection:
collection = [collection]
else:
collection = doc.get("collection_names", [])
collection = set(collection).difference(extracted_collections)
if not collection:
log.debug(f"skipping {doc} as it has already been extracted")
continue
try:
if doc["type"] == "text":
context = doc["content"]
elif doc["type"] == "collection":
context = query_embeddings_collection(
collection_names=doc["collection_names"],
query=query,
k=k,
r=r,
embeddings_function=embeddings_function,
reranking_function=reranking_function,
)
@ -252,22 +269,26 @@ def rag_messages(
collection_name=doc["collection_name"],
query=query,
k=k,
r=r,
embeddings_function=embeddings_function,
reranking_function=reranking_function,
)
except Exception as e:
log.exception(e)
context = None
if context:
relevant_contexts.append(context)
extracted_collections.extend(collection)
log.debug(f"relevant_contexts: {relevant_contexts}")
context_string = ""
for context in relevant_contexts:
if context:
context_string += " ".join(context["documents"][0]) + "\n"
items = context["documents"][0]
context_string += "\n\n".join(items)
context_string = context_string.strip()
ra_content = rag_template(
template=template,
@ -275,6 +296,8 @@ def rag_messages(
query=query,
)
log.debug(f"ra_content: {ra_content}")
if content_type == "list":
new_content = []
for content_item in user_message["content"]:
@ -321,15 +344,14 @@ def generate_openai_embeddings(
from typing import Any
from langchain_core.callbacks import CallbackManagerForRetrieverRun
from langchain_core.documents import Document
from langchain_core.retrievers import BaseRetriever
from langchain_core.callbacks import CallbackManagerForRetrieverRun
class ChromaRetriever(BaseRetriever):
collection: Any
k: int
embeddings_function: Any
top_n: int
def _get_relevant_documents(
self,
@ -341,7 +363,7 @@ class ChromaRetriever(BaseRetriever):
results = self.collection.query(
query_embeddings=[query_embeddings],
n_results=self.k,
n_results=self.top_n,
)
ids = results["ids"][0]
@ -355,3 +377,60 @@ class ChromaRetriever(BaseRetriever):
)
for idx in range(len(ids))
]
import operator
from typing import Optional, Sequence
from langchain_core.documents import BaseDocumentCompressor, Document
from langchain_core.callbacks import Callbacks
from langchain_core.pydantic_v1 import Extra
from sentence_transformers import util
class RerankCompressor(BaseDocumentCompressor):
embeddings_function: Any
reranking_function: Any
r_score: float
top_n: int
class Config:
extra = Extra.forbid
arbitrary_types_allowed = True
def compress_documents(
self,
documents: Sequence[Document],
query: str,
callbacks: Optional[Callbacks] = None,
) -> Sequence[Document]:
if self.reranking_function:
scores = self.reranking_function.predict(
[(query, doc.page_content) for doc in documents]
)
else:
query_embedding = self.embeddings_function(query)
document_embedding = self.embeddings_function(
[doc.page_content for doc in documents]
)
scores = util.cos_sim(query_embedding, document_embedding)[0]
docs_with_scores = list(zip(documents, scores.tolist()))
if self.r_score:
docs_with_scores = [
(d, s) for d, s in docs_with_scores if s >= self.r_score
]
result = sorted(docs_with_scores, key=operator.itemgetter(1), reverse=True)
final_results = []
for doc, doc_score in result[: self.top_n]:
metadata = doc.metadata
metadata["score"] = doc_score
doc = Document(
page_content=doc.page_content,
metadata=metadata,
)
final_results.append(doc)
return final_results

19
backend/config.py
View file

@ -420,6 +420,9 @@ if WEBUI_AUTH and WEBUI_SECRET_KEY == "":
CHROMA_DATA_PATH = f"{DATA_DIR}/vector_db"
# this uses the model defined in the Dockerfile ENV variable. If you dont use docker or docker based deployments such as k8s, the default embedding model will be used (sentence-transformers/all-MiniLM-L6-v2)
RAG_TOP_K = int(os.environ.get("RAG_TOP_K", "5"))
RAG_RELEVANCE_THRESHOLD = float(os.environ.get("RAG_RELEVANCE_THRESHOLD", "0.0"))
RAG_EMBEDDING_ENGINE = os.environ.get("RAG_EMBEDDING_ENGINE", "")
RAG_EMBEDDING_MODEL = os.environ.get(
@ -431,10 +434,9 @@ RAG_EMBEDDING_MODEL_TRUST_REMOTE_CODE = (
os.environ.get("RAG_EMBEDDING_MODEL_TRUST_REMOTE_CODE", "").lower() == "true"
)
RAG_RERANKING_MODEL = os.environ.get(
"RAG_RERANKING_MODEL", "BAAI/bge-reranker-v2-m3"
)
log.info(f"Reranking model set: {RAG_RERANKING_MODEL}"),
RAG_RERANKING_MODEL = os.environ.get("RAG_RERANKING_MODEL", "")
if not RAG_RERANKING_MODEL == "":
log.info(f"Reranking model set: {RAG_RERANKING_MODEL}"),
RAG_RERANKING_MODEL_TRUST_REMOTE_CODE = (
os.environ.get("RAG_RERANKING_MODEL_TRUST_REMOTE_CODE", "").lower() == "true"
@ -448,16 +450,15 @@ if USE_CUDA.lower() == "true":
else:
DEVICE_TYPE = "cpu"
CHROMA_CLIENT = chromadb.PersistentClient(
path=CHROMA_DATA_PATH,
settings=Settings(allow_reset=True, anonymized_telemetry=False),
)
CHUNK_SIZE = 1500
CHUNK_OVERLAP = 100
CHUNK_SIZE = int(os.environ.get("CHUNK_SIZE", "1500"))
CHUNK_OVERLAP = int(os.environ.get("CHUNK_OVERLAP", "100"))
RAG_TEMPLATE = """Use the following context as your learned knowledge, inside <context></context> XML tags.
DEFAULT_RAG_TEMPLATE = """Use the following context as your learned knowledge, inside <context></context> XML tags.
<context>
[context]
</context>
@ -471,6 +472,8 @@ And answer according to the language of the user's question.
Given the context information, answer the query.
Query: [query]"""
RAG_TEMPLATE = os.environ.get("RAG_TEMPLATE", DEFAULT_RAG_TEMPLATE)
RAG_OPENAI_API_BASE_URL = os.getenv("RAG_OPENAI_API_BASE_URL", OPENAI_API_BASE_URL)
RAG_OPENAI_API_KEY = os.getenv("RAG_OPENAI_API_KEY", OPENAI_API_KEY)

5
backend/main.py
View file

@ -120,12 +120,13 @@ class RAGMiddleware(BaseHTTPMiddleware):
data["messages"],
rag_app.state.RAG_TEMPLATE,
rag_app.state.TOP_K,
rag_app.state.RELEVANCE_THRESHOLD,
rag_app.state.RAG_EMBEDDING_ENGINE,
rag_app.state.RAG_EMBEDDING_MODEL,
rag_app.state.sentence_transformer_ef,
rag_app.state.sentence_transformer_rf,
rag_app.state.RAG_OPENAI_API_KEY,
rag_app.state.RAG_OPENAI_API_BASE_URL,
rag_app.state.OPENAI_API_KEY,
rag_app.state.OPENAI_API_BASE_URL,
)
del data["docs"]

1
src/lib/apis/rag/index.ts
View file

@ -123,6 +123,7 @@ export const getQuerySettings = async (token: string) => {
type QuerySettings = {
k: number | null;
r: number | null;
template: string | null;
};

2
src/lib/components/chat/Settings/Chats.svelte
View file

@ -2,7 +2,7 @@
import fileSaver from 'file-saver';
const { saveAs } = fileSaver;
import { chats, user } from '$lib/stores';
import { config, chats, user } from '$lib/stores';
import {
createNewChat,

29
src/lib/components/documents/Settings/General.svelte
View file

@ -42,6 +42,7 @@
let querySettings = {
template: '',
r: 0.0,
k: 4
};
@ -124,7 +125,7 @@
updateRerankingModelLoading = true;
const res = await updateRerankingConfig(localStorage.token, {
reranking_model: rerankingModel,
reranking_model: rerankingModel
}).catch(async (error) => {
toast.error(error);
await setRerankingConfig();
@ -450,6 +451,12 @@
</div>
</div>
<div class="mt-2 mb-1 text-xs text-gray-400 dark:text-gray-500">
{$i18n.t(
'Note: If you choose a reranking model, it will use that to score and rerank instead of the embedding model.'
)}
</div>
<hr class=" dark:border-gray-700 my-3" />
<div class=" flex w-full justify-between">
@ -576,6 +583,26 @@
</div>
</div>
<div class=" flex">
<div class=" flex w-full justify-between">
<div class="self-center text-xs font-medium flex-1">
{$i18n.t('Relevance Threshold')}
</div>
<div class="self-center p-3">
<input
class=" w-full rounded-lg py-1.5 px-4 text-sm dark:text-gray-300 dark:bg-gray-850 outline-none"
type="number"
step="0.01"
placeholder={$i18n.t('Enter Relevance Threshold')}
bind:value={querySettings.r}
autocomplete="off"
min="0.0"
/>
</div>
</div>
</div>
<div>
<div class=" mb-2.5 text-sm font-medium">{$i18n.t('RAG Template')}</div>
<textarea