HomeArtificial IntelligenceVector databases: Shiny object syndrome and the case of a missing unicorn

Vector databases: Shiny object syndrome and the case of a missing unicorn

Welcome to 2024. If you don't ride the generative AI wave, you may as well be stuck in 2022 – effectively ancient history within the AI ​​timeline. Every organization now has an AI roadmap, from AI pillows to AI toothbrushesand when you still haven't put together a plan quickly, let me suggest a three-step roadmap.

Step 1: Gather a team to finish the duty Andrew Ng Of course, because nothing is as relevant as a certificate of completion.

Step 2: Get the API keys from OpenAI. No, you may't access ChatGPT, it's not a thing.

Step 3: Vector databaseEmbeds, technical wizardry!

Now the show can begin: Put all the information into the vector database and add somewhat RAG architecture, sprinkle in a little bit of quick technique and voilà! The Gen AI wave has officially arrived in your organization. Now sit back, chill out and revel in the exciting wait for the magic to occur. Waiting, waiting…still waiting. Ah, the sweet anticipation of the awesomeness of genetic AI!

In the chaotic sprint to genetic AI and its seemingly uncomplicated implementation large language model (LLM) Architecture hiccups occur when firms forget use cases and begin chasing the technology. When AI is your hammer, every problem seems solvable.

And while LLMs and vector databases appear to be in vogue (Taylor Swift is more trendy), the notion of vector-based representations, crucial to modern natural language processing, has deep roots.

Word Associations: Review of “Who Wants a Million Dollars?”

George Miller's book , published in 1951 and derived from his earlier works, expands the concept of distributional semantics. Miller suggested that words appearing in similar contexts are more likely to have similar meanings, laying the muse for vector-based representations.

He further showed that associations between words have strengths by stating, “At a more molecular level, the strength of 'I' appears to differ greatly from moment to moment.” It is a most unlikely answer to the query, “Who was that?” first king of England?” and a really likely response to “Who wants one million dollars?” While a dog can evoke an associative response to “animal,” the association from “animal” to “dog” is weak, as Miller concluded: “The association, like a vector, has each magnitude and direction.”

Word associations return even further, as a study by shows Kent and Rosanoflf Participants were asked “the primary word that involves mind apart from the stimulus word.”

Thomas K. Landauer's work, “A Solution to Plato's Problem: The Latent Semantic Analysis Theory of the Acquisition, Induction, and Representation of KnowledgePublished in 1997, addresses the main points of vector-based representation of concepts. Introduced by Landauer, latent semantic evaluation (LSA) uses mathematical techniques similar to singular value decomposition to create vector spaces by which words with similar meanings are positioned close to one another. This facilitates the efficient calculation of semantic relationships and contributes to tasks similar to information retrieval and document categorization.

In 2003, Yoshua Bengio, Réjean Ducharme and Pascal Vincent published “A neural probabilistic language model“, introducing a neural network model that may learn word embeddings. This paper marked a notable shift toward neural network-based approaches to word representation and laid the muse for word2vec, GloVe, ELMO, BERT, and the present suite of embedding models.

Vector-based text representations are nothing latest and have continued to evolve, but when does the Vector DB show begin?

When does the Vector DB show start?

The Vector DB space is becoming increasingly crowded and each vendor is striving to face out from the plethora of features. Performance, scalability, ease of use, and pre-built integrations are only a number of the aspects that shape their differentiation. The crux of the matter, nonetheless, lies in relevance – it’s all the time higher to get the appropriate answer in a number of seconds and even minutes than to get the flawed answer in a flash.

When we delve into the intricacies of strict vector search (never a superb idea, see below), the pivot point is the approximate nearest neighbor (ANN). Vector DBs offer quite a lot of ANNs, each with its own variant:

As the terms and details turn into fuzzy, the seemingly straightforward LLM architecture not appears easy. However, when you had the alternative to generate embeds of your data using OpenAI APIs and retrieve them using the identical ANNs as HSNW, wouldn't the relevance (or irrelevance) be the identical?

“Can you fix my computer?” No, but I can let you know that bananas are berries and strawberries usually are not.

Let's have a look at how someone might use the system and whether it really is sensible to convert the information into vectors. Imagine this scenario: A user enters a straightforward query like “Error 221” to search out manuals that may help resolve it. We do the same old thing: convert the query to its embedding, retrieve it using a variation of ANN, and rating it using cosine similarity. Standard material, right? The highlight: The results mean that a document with error 222 receives the next rating than the one with error 221.

Yes, it's like saying, “Find error 221,” and the system says, “Here's something about error 222; I hope this helps!” Not exactly what the user signed up for. So let's not only dive headfirst into the world of vectors without determining if it's the appropriate move.

Hype aside, what's happening?

What's flawed with vector databases anyway? It's all about information retrieval, but let's be honest, that's nothing latest, although it would feel prefer it with all of the hype around it. We have already got SQL databases, NoSQL databases, full-text search apps and vector libraries that do that job. Sure, vector databases offer semantic retrieval, which is great, but SQL databases prefer it Single store and Postgres (with the pgvector Extension) also can handle semantic retrieval while providing standard DB functions similar to ACID. Full-text search applications like Apache Solr, Elasticsearch And OpenSearch are also rocking the vector search scene, together with search products like Coveoand convey something with you serious word processing Hybrid search capabilities.

But that's the thing about vector databases: they're form of stuck in the center. They cannot completely replace traditional databases and still lag behind in supporting the word processing capabilities required for comprehensive search capabilities. The Dragon considers hybrid search to be just attribute filtering using Boolean expressions!

“If technology isn’t your differentiator, go for hype.”

Pinecones Hybrid Search brings with it each a warning and caveats, and although some may argue that it did ahead of his timeIf the celebrations had to attend for the OpenAI revolution a number of years later, attending to the party early wouldn't matter much.

It wasn’t that early either – Weave, Vespa And Mivlus were already available on the market with their Vector DB offerings, and Elasticsearch, OpenSearch and Solr were ready around the identical time. If technology isn't your unique selling point, go for hype. Pine cones $100 million Series B funding was directed by Andreessen Horowitz, who in some ways lives by the script for which he was created the boom times in technology. And with all of the hype surrounding the AI ​​revolution and Gen AI, the Gen AI corporate party still hasn't began. Time will tell whether Pinecone is a lost unicorn, but distinguishing it from other vector databases can be an increasing challenge.

Shiny object syndrome

Pursue the search is difficult. Rarely is the answer to easily load data into vector storage and expect miracles to occur. From splitting PDFs to the appropriate size to organising the appropriate access controls, all the pieces requires careful planning and execution to make sure optimal performance and value. If your organization's use case is to look a limited variety of documents, scalability is probably not a pressing issue. If your use case is heavily focused on keyword searches, as shown in Figure 3, jumping into vector implementation also can backfire.

Ultimately, it doesn't matter to the top user whether it's a vector search, a keyword search, a rule-driven search, or perhaps a “phone a friend” search. The most vital thing for the user is to get the right answer. This rarely happens by relying solely on one methodology. Understand your use case and validate your test scenarios… and… don’t be lured by shiny objects simply because they’re popular.


Please enter your comment!
Please enter your name here

Must Read