Building a semantic search experience using PHP and Meilisearch 

Hello there 

Hellcome Guillaume Loulier / @Guikingone Break things @SensioLabs Work mainly with PHP and Rust, a bit of Zig (thanks to ZML) Publish a weekly newsletter about cloud computing, machine learning, PHP and more on Substack 

Summary 01 - What do you mean? 02 - Are computers capable of understanding? 03 - What about a demo? 04 - Any advices? 01 

But really, what do you mean? 

Ego time 

Searching is hard Until recently, search experiences / engines were built on top of keyword search Keyword search is based on sentence occurrences, hits and a bit of luck What if a sentence contains many times the word we’re searching for? Is it relevant? 

Everything has a meaning The human brain is wired to perform “logical thinking” via memories and patterns, computers are tied to binaries and mathematics Trying to search in a “logical way” via computers is a path to failure Could mathematics helps us? How? 

Did you said maths? But, wait, what are words and sentences if not sequences of numbers in the alphabet? What if we could expose words into a three dimensional space and see them? What about images? If everything is a number, could we perform calculations around them and find patterns? 

But wait, what about “IA”? 

There’s no such thing as IA Intelligence is not “general”, think of it as a subset of specific intelligences What we call “IA” is just a buzzword for algorithms and workflows At the end, everything comes down to energy transformation and this topic is quite complex (and for another day) 

Understanding the world The recent revolutions of machine learning are LLMs and transformers, a subset of neural networks The concept behind is understanding and dealing with context LLMs are not intelligent, they excel at probabilities and tokenization, at the end, they’re all biased 

Similarity is about differences 

Maths, vectors and friends Semantic search (aka vector search) is about proximity and similarity If we transform everything to numbers, we can easily perform operations on them The closer numbers are, the more similar the content seems to be, the more relevant it seems 

Sadly Similarity doesn’t mean relevancy neither intention 0.1 is close to 0.2 but also to 0, wait, what’s behind 0? Depending on the context, similarity can introduce bias Context is key, everything else is fog The more context you gave, the better the results will be Similarity doesn’t mean that the meaning is correct “Hello world” is similar to “Hi there” but “Hello sir” is more relevant 

Visualizing the invisible 

Context is meaningful The dog is running after the tennis ball in a kitchen The dog is running after the tennis ball in a kitchen The tennis ball is running after the dog in a kitchen 

Let’s vectorize it* The dog is running after the tennis ball in a kitchen 0 -1 0 1 -1 0 1 1 1 0 -1 0 2 2 2 0 0 2 -1 0 2 2 3 0 0 2 3 3 0 1 0 -1 2 1 1 2 0 1 -1 3 2 1 0 1 Encoder [-0.88440161, -0.00996133, 0.243678553, …] * Values are built for example 

Meilisearch 

Becoming speed Build for a specific use case at LVMH First iteration built in Go, rewritten in Rust, speaking of speed (< 50ms) Open-source (MIT) and built in France PHP SDK, Symfony (Sylius?) bundle and many more 

Tease me Built-in vector store Can be used with Ollama, OpenAI, Anthropic APIs and more Unified entrypoint for keyword and vector search Fine-tuning available, multi / federated search compatible 

Show me 

Force of habits Built on top of symfony/demo Available via https://github.com/Guikingone/SyliusconSemantic 

I was there, a long time ago… 

Be contextful The more precise you are, the easier it will be to retrieve (seems obvious but…) Choosing the right model is hard, just experiment and/or fine-tune them Want to go further? What about RAG? At the end of the day, everything is tied to mathematics, don’t reinvent the wheel 

Oh, by the way… Think about user interactions rather than hits / results Semantic search is AN idea, not A promise Sometimes, a plain-old keyword search is enough and that’s fine Experiment, fine-tune, improve, learn and monitor 

Questions?