Try SuperlinkedRetrieval-Augmented Generation (RAG) is revolutionizing traditional search engines and AI methodologies for information retrieval. However, standard RAG systems employing simple semantic search often lack efficiency and precision when dealing with extensive data repositories. Hybrid search, on the other hand, combines the strengths of different search methods, unlocking new levels of efficiency and accuracy. Hybrid search is flexible and can be adapted to tackle a wider range of information needs.
Hybrid search can also be paired with semantic reranking (to reorder outcomes) to further enhance performance. Combining hybrid search with reranking holds immense potential for various applications, including natural language processing tasks like question answering and text summarization, even for implementation at a large-scale.
In our article, we'll delve into the nuances and limitations of hybrid search and reranking. Though pure vector search is preferable, in many cases hybrid search can enhance the retrieval component in RAG (Retrieval Augmented Generation), and thereby deliver impactful and insightful text generation across various domains.
In current Retrieval-Augmented Generation (RAG) systems, word embeddings are used to represent data in the vector database, and vector similarity search is commonly used for searching through them. For LLMs and RAG systems, embeddings - because they capture semantic relationships between words - are generally preferred over keyword-based representations like Bag-of-words (BoW) approaches.
But each of vector similarity search and keyword search has its own strengths and weaknesses. Vector similarity search is good, for example, at dealing with queries that contain typos, which usually don’t change the overall intent of the sentence. However, vector similarity search is not as good at precise matching on keywords, abbreviations, and names, which can get lost in vector embeddings along with the surrounding words. Here, keyword search performs better.
That being said, keyword search is not as good as vector similarity search at fetching relevant results based on semantic relationships or meaning, which are only available via word embeddings. For example, a keyword search will relate the words “the river bank” and “the Bank of America” even though there is no actual semantic connection between the terms - a difference to which vector similarity search is sensitive. Keyword search would, therefore, benefit from vector search, but the prevailing approach is not to combine them but rather to implement them separately using distinct methodologies.
**In hybrid search - a keyword-sensitive semantic search approach, we combine vector search and keyword search algorithms to **take advantage of their respective strengths while mitigating their respective limitations.
Vector similarity search proves to be inadequate in certain scenarios, including:
In such instances, hybrid search is extremely useful. Keyword search guarantees that abbreviations, annotations, names, and code stay on the radar, while vector search ensures that search results are contextually relevant. The new Stack Overflow, for example, has adopted just such an approach.
Stack Overflow has moved from simple lexical search to semantic and hybrid search, leveraging AI to make its search algorithms more powerful. Their earlier approach, which used the TF-IDF algorithm to match on user search string keywords - including exact code snippets, could miss posts that were relevant but didn't contain those exact keywords. Stack Overflow's new hybrid search methodology finds semantically relevant content from their corpus, and also matches the exact code entered by the user if necessary. This significantly improved Stack Overflow's search results.
While hybrid search confers advantages in many use cases, it is not a silver bullet. It has limitations, including:
That being said, there are many vector databases that incorporate functions that implement hybrid search - e.g., Pinecone, ElasticSearch, Apache Cassandra, and Weaviate. Check out the Vector DB Comparison table to see which vector databases support hybrid search.
The hybrid search algorithm combines keyword search and vector search to retrieve relevant content from a corpus. Let's take a look at the components that make up the architecture of hybrid search.
Sparse vectors are vectors with high dimensionality, where most elements are zero. They usually symbolize various language tokens, non-zero values signifying their respective importance. Keyword search is also called sparse vector search. The BM25 (Best Match 25) algorithm is a popular and effective ranking function employed for keyword matching in embeddings. BM25 finds the most relevant documents for a given query by examining two things:
The BM25 score for document D for query Q is calculated as the sum of the scores for individual query terms. Here's the formula for calculating the BM25 score:
BM25(D, Q) = ∑(IDF(q) * ((TF(q, D) * (k1 + 1)) / (TF(q, D) + k1 * (1 — b + b * (|D| / avgdl)))))
where,
Notice that the BM25 algorithm is a refined version of the TF-IDF(Term-Frequency Inverse-Document Frequency) algorithm.
Dense vectors, or embeddings are arrays with a high number of dimensions, filled predominantly with meaningful, non-zero values. Machine learning frequently employs these to represent the underlying semantics and connections of words in a numerical format, thereby effectively encapsulating their semantic essence. Dense vector search is a method used in semantic search systems for finding similar items in a vector space.
A common approach to vector search is cosine similarity search. Cosine similarity is calculated as the result of the dot product of the vectors, normalized by the multiplication of their magnitudes. The nearer the outcome is to 1, the more similar the vectors are.
C(A,B) = cos(θ) = A.B / ||A|| ||B||
The results from each algorithm have to be fused to implement a hybrid search. There are various strategies to combine them and get a score. To balance the keyword search score and vector search score to meet our requirements, we use the following formula:
H = (1-α) K + αV
where,
The hybrid score is a pure vector score when α is 1, and a pure keyword score when α is 0.
Reciprocal Rank Fusion (RRF) is one of several available methods for combining dense and sparse search scores. RRF ranks each passage according to its place in the keyword and vector outcome lists, and then merges these rankings to generate a unified result list. The RRF score is determined by summing the inverse rankings from each list. Positioning the document’s rank in the denominator imposes a penalty on documents that appear lower in the list.
where,
Typically, algorithms yield top-k matches. But these top-k matches may not always include the relevant sections, or, conversely, not all relevant sections may be within these top-k matches. We can ameliorate this issue by ranking all retrieved content based on a score indicating semantic relevance to the query.
To do this, responses from the retriever must be passed to a semantic scoring model. Semantic scoring models are transformer models that take in queries and documents to produce a score in a calibrated range. This reranking process returns a list of documents, sorted according to relevance score, from highest to lowest, and incorporated into the response payload of the query.
Let’s test the performance of a normal vector search algorithm and a hybrid search algorithm in various contexts. We will be using ChromaDB supported by LangChain and models from HuggingFace. ChromaDB has no direct implementations for hybrid search, but for clarity, we will create an ensemble in the same way we discussed in the theory.
First, we install and import the required libraries.
!pip install langchain langchain-community rank_bm25 pypdf unstructured chromadb !pip install unstructured['pdf'] unstructured !apt-get install poppler-utils !apt-get install -y tesseract-ocr !apt-get install -y libtesseract-dev !pip install pytesseract !pip install bitsandbytes accelerate peft safetensors sentencepiece
Next, we load our document and split it into chunks of the required size.
from langchain.document_loaders import UnstructuredPDFLoader from langchain.text_splitter import RecursiveCharacterTextSplitter from langchain.vectorstores import Chroma from langchain.embeddings import HuggingFaceInferenceAPIEmbeddings from langchain.llms import HuggingFaceHub import torch from transformers import ( AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, pipeline, ) from langchain import HuggingFacePipeline from langchain.retrievers import BM25Retriever, EnsembleRetriever import os
Now, we load the PDF document and split it into chunks of the desired length with sufficient overlap. In this step, you can adjust the chunk size based on the length of your document and the requirements of your LLM.
doc_path = "/content/document.pdf" file = UnstructuredPDFLoader(doc_path) docs = file.load()
# create chunks splitter = RecursiveCharacterTextSplitter(chunk_size=200, chunk_overlap=30) chunks = splitter.split_documents(docs)
Next, we create a vector store using the embeddings we obtain from the text.
embeddings = HuggingFaceInferenceAPIEmbeddings( api_key=HF_TOKEN, model_name="BAAI/bge-base-en-v1.5" ) vectorstore = Chroma.from_documents(chunks, embeddings)
Now, we build the keyword and semantic retrievers separately. For keyword matching, we use the BM25 retriever from Langchain. By setting k to 3, we’re asking the retriever to return the 3 most relevant documents or vectors from the vector store.
vectorstore_retreiver = vectorstore.as_retriever(search_kwargs={"k": 3}) keyword_retriever = BM25Retriever.from_documents(chunks) keyword_retriever.k = 3
Now, we create the ensemble retriever, which is a weighted combination of the keyword and semantic retrievers above.
ensemble_retriever = EnsembleRetriever(retrievers=[vectorstore_retreiver, keyword_retriever], weights=[0.3, 0.7])
We can modify the weights parameter to balance the impact of both search outcomes appropriately as needed. The weight values correspond to α and 1-α, as we discussed above. Here, we have weighted keywords more heavily, with a value of 0.7.
Our RAG pipeline needs an LLM. We utilize a quantized version of Zephyr-7B-Beta for lightweight and optimized performance.
model_name = "HuggingFaceH4/zephyr-7b-beta" # function for loading 4-bit quantized model def load_quantized_model(model_name: str): """ model_name: Name or path of the model to be loaded. return: Loaded quantized model. """ bnb_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_use_double_quant=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16, ) model = AutoModelForCausalLM.from_pretrained( model_name, load_in_4bit=True, torch_dtype=torch.bfloat16, quantization_config=bnb_config, ) return model # initializing tokenizer def initialize_tokenizer(model_name: str): """ model_name: Name or path of the model for tokenizer initialization. return: Initialized tokenizer. """ tokenizer = AutoTokenizer.from_pretrained(model_name, return_token_type_ids=False) tokenizer.bos_token_id = 1 # Set beginning of sentence token id return tokenizer tokenizer = initialize_tokenizer(model_name) model = load_quantized_model(model_name) # specify stop token ids stop_token_ids = [0] # build huggingface pipeline for using zephyr-7b-beta pipeline = pipeline( "text-generation", model=model, tokenizer=tokenizer, use_cache=True, device_map="auto", max_length=2048, do_sample=True, top_k=5, num_return_sequences=1, eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.pad_token_id, ) llm = HuggingFacePipeline(pipeline=pipeline)
Now, we define the hybrid search and semantic search retrievers.
from langchain.chains import RetrievalQA normal_chain = RetrievalQA.from_chain_type( llm=llm, chain_type="stuff", retriever=vectorstore_retreiver ) hybrid_chain = RetrievalQA.from_chain_type( llm=llm, chain_type="stuff", retriever=ensemble_retriever ) response = PREFFERED_CHAIN.invoke("QUERY")
Let’s check responses from both retrievers in various contexts. First, we will query in a general context without keywords, abbreviations, or location filters.
Query: What are the two strategic challenges that the United States faces according to the National Security Strategy? Semantic Search Result: According to the given context, the United States faces two strategic challenges as stated in the 2022 National Security Strategy. The challenges are that the post-Cold War era is over, and a competition is underway between major powers to shape what comes next. Hybrid Search Result: According to the 2022 National Security Strategy, the United States faces two strategic challenges: the end of the post-Cold War era and a competition underway between major powers to shape what comes next, and the recognition that the democratic allies and partners in the Indo-Pacific and Europe are mutually reinforcing and intertwined fates.
Both results are somewhat accurate and satisfactory. Now, let us try again, this time incorporating abbreviations and name filters to see how that works.
Query: How did Joe Biden help deepen NATO relationship? Semantic Search Result: Under Joe Biden's leadership, America has deepened its relationship with NATO by responding to global challenges and strengthening NATO, making it stronger and more united than it has ever been, as we look to continue promoting stability, prosperity, and growth. Hybrid Search Result: Under Joe Biden's leadership, the United States has worked to deepen NATO's relationship through various means. First, he has emphasized the importance of NATO's collective defense commitment, reiterating that the United States will defend every inch of NATO territory, which has helped build trust and confidence among alliance members. Second, Biden has played a key role in modernizing NATO's strategies and capabilities to address emerging threats, such as cyber-attacks and hybrid warfare. Third, he has prioritized engagement with key NATO partners, such as Sweden and Finland, to broaden and deepen the transatlantic bond. Fourth, Biden has advocated for increased funding and resources for NATO, which has helped strengthen the alliance's military and economic capabilities. Overall, Biden's efforts have contributed to a more cohesive and effective NATO, which is better equipped to address global challenges and opportunities.
The results of the hybrid search are more detailed, provide a more comprehensive response, including the names of countries and other specific information. Now, let’s add specific locations to see how each approach fares.
Query: How is Strait of Hormuz important in Biden's strategy? Semantic Search Result: The passage doesn't explicitly state how the Strait of Hormuz is important in Biden's strategy. However, it does mention that the United States and its allies have a shared responsibility to maintain a free and open Indo-Pacific, which includes protecting the freedom of navigation. It is possible that preserving freedom of navigation in this area is also a part of Biden's overall strategy for the region. Hybrid Search Result: Biden's strategy emphasizes the importance of ensuring freedom of navigation through the Strait of Hormuz, a waterway located in the Middle East. This strategy aims to prevent any country from dominating the region through military efforts and ensures that there are no efforts to control the waterways. This emphasis on freedom of navigation is crucial for the United States and its allies as a significant portion of the world's oil supply passes through this waterway. Any disruption or control of this waterway could have significant economic and geopolitical implications, making Biden's strategy to maintain this freedom critical.
The hybrid search appears to perform better in providing a specific and detailed response to the query, whereas the semantic search produced a more generalized interpretation without explicitly addressing the importance of the Strait of Hormuz and a geographical overview of the place.
Our implementation example above uses ChromaDB. Your use case may warrant using a different database. Other databases, for example, Weaviate DB, offer native support and implementation for hybrid search. Here's how you would define the retriever component for hybrid search in Weaviate DB.
from langchain.retrievers.weaviate_hybrid_search import WeaviateHybridSearchRetriever retriever = WeaviateHybridSearchRetriever( alpha = 0.5, # defaults to 0.5, which is equal weighting between keyword and semantic search client = client, # keyword arguments to pass to the Weaviate client index_name = "", # The name of the index to use text_key = "", # The name of the text key to use attributes = [], # The attributes to return in the results ) hybrid_chain = RetrievalQA.from_chain_type( llm=llm, chain_type="stuff", retriever=retriever )
The value of the alpha parameter in the Weaviate retriever can be adjusted to control the relative impact of semantic and keyword searches.
Because the retrievers created above score the top k responses internally and return the highest-scoring response, we may not always need to perform reranking explicitly. In the event of low accuracy in the retrieved content, you can implement a reranker directly using libraries from Cohere, or build your own custom reranking function. When using a reranker from Cohere, the following changes should be made in the retriever.
from langchain.retrievers import ContextualCompressionRetriever from langchain.retrievers.document_compressors import CohereRerank compressor = CohereRerank() compression_retriever = ContextualCompressionRetriever( base_compressor=compressor, base_retriever=ensemble_retriever ) hybrid_chain = RetrievalQA.from_chain_type( llm=llm, chain_type="stuff", retriever=compression_retriever )
We've looked at how RAG system performance can be enhanced by using hybrid search along with reranking, as compared with using keyword search or vector search on their own. By combining keyword and vector search into one hybrid method, we can match on keywords in contextually relevant content, achieving more refined responses. Using hybrid search, the retriever's higher recall rates permit the LLM to produce higher-quality outputs.
Stay updated with VectorHub
Continue Reading
