The Improviser · Lesson 7 — Cutting the Tape ← Course

Cutting the tape — chunking.

You don't hand the soloist the whole session. You cut it into takes, and pull only the one that matters.

The library isn't shelved as whole sessions. Up front, every recording is cut into small passages called chunks, and each chunk is embedded and indexed on its own — so a search returns single passages, never a whole recording.

Why? A full session log might run to hundreds of lines. If the search could only return the entire log, you would blow the context budget (Lesson 2) and bury the one relevant take in noise. Cut it into chunks, and the search can hand back just the passage you need.

You wouldn't reel through the whole night to cue up one tune. You cut the tape into takes — one per number — and call up only the take that matters. Same idea.

Chunk size is a judgement call: too big and each piece is unfocused and wasteful; too small and it loses the context around it. A paragraph, or a few sentences, is the usual sweet spot.

Pick what you're after and watch the search pull just the one take it needs out of the whole log — and see how much it left behind. Tap a take to hear it.

One session log, cut into five takes. Pick what you need and see which single take is pulled.
the session log · cut into chunks

Cut once, retrieve the piece.

Splitting happens up front, when the recording is added. At question time the search returns just the matching chunk — not the whole log. 


      

    


    

      
Small takes, not whole reels.

      

        Recordings are split into chunks so the search can return just the passage that
        answers the question — focused, and small enough to fit the window. You retrieve the
        take, not the whole night.
      

      

        Chunk size is a trade-off: big enough to keep the surrounding context, small enough to
        stay focused and fit the budget. Get it right and every retrieval is a clean, relevant
        take.
      

      

        Next: the wrong clip →
        
      

    


  


  

    

      ▸ Go deeper — getting chunking right optional
      

        
Why not just retrieve the whole recording

        

          Two reasons. It costs budget — a giant log eats the window (Lesson 2). And it buries
          the signal — the one relevant take sits in a sea of irrelevant ones, and the model
          attends worse to a big blob than to a tight, on-point passage.
        

        
Size and overlap

        

          Too big and a chunk is unfocused; too small and it loses the context a detail needs to
          make sense. Chunks often overlap a little at their edges, so a phrase split
          across a boundary is not lost from both halves.
        

        
Cut on natural boundaries

        

          Splitting blindly every N characters can slice a sentence in half. Better to cut on
          structure — takes, sections, headings — so each chunk is one coherent thought that
          stands on its own.
        

        
Tag each chunk

        

          Store metadata alongside each chunk — which recording it came from, the set, the date.
          That is what lets the system cite a source, filter by recency, and tell you exactly
          where an answer came from.