The Video Assistant · Lesson 7 — By Meaning ← Course

By meaning — embeddings & semantic search.

How the search finds the right clips even when the words don't match.

Page 1 said the search finds clips “by meaning”. Here is how. Every piece of text — each clip, and your question — is turned into a position in a space of meaning. Similar meanings sit close together; different meanings sit far apart. That position is called an embedding.

The positions are worked out from meaning, not spelling. “Flagged offside” and “beyond the last defender” share no words, yet they land right next to each other, because they describe the same kind of event. “Handball shout” lands in a different neighbourhood, and “substitution” further off still.

So searching becomes geometry. Drop your question into the same space and grab its nearest neighbours — the closest few points. No keyword matching; just “what sits closest in meaning”.

In reality the space is not two dimensions but hundreds or thousands — each embedding is a long list of numbers. But the idea is exactly the picture below: near means similar, far means different.

Pick a question and watch it drop onto the meaning map. The nearest clips light up — even the ones that share none of the same words.

Each dot is a clip, placed by meaning. Pick a question and see which sit closest.
meaning map · near = similar

Turn text into a point, then find the nearest.

Embedding is done once and stored. At question time you embed the question and take its closest neighbours — that is the whole of “search by meaning”. 


      

    


    

      
Near in meaning, not in words.

      

        An embedding is a position that captures what text means, so similar ideas
        sit close together no matter what words they use. Search then becomes geometry: turn
        the question into a point and take its nearest neighbours.
      

      

        This is why RAG can find the right record even when your question and the record
        share not a single word — it is matching meaning, not spelling.
      

      

        Next: the same step →
        
      

    


  


  

    

      ▸ Go deeper — what an embedding really is optional
      

        
It's a long list of numbers

        

          The 2D map here is a flattened picture. A real embedding is a vector of hundreds or
          thousands of numbers — a position in a very high-dimensional space. Closeness is
          measured with a distance or, more often, the angle between the two vectors (cosine
          similarity — the maths, if you want it later).
        

        
Where the positions come from

        

          A separate embedding model, trained on huge amounts of text, learned to place
          text so that similar meanings land nearby. You run each document through it once, up
          front, and store the results — so the heavy lifting happens before any question is
          ever asked.
        

        
Same space for both

        

          The question and the documents have to be embedded by the same model, or their
          positions are not comparable — like plotting two maps on different scales and trying
          to measure the distance between them.
        

        
Why keyword search isn't enough

        

          Plain keyword search misses synonyms and paraphrases — “offside” never matches
          “last defender”. Semantic search catches them, because it works on meaning. Many real systems
          run both and blend the results, which is called hybrid search.