Where we left off

Lesson 4 turned an image into a caption, plain text. This lesson does nothing new conceptually, it hands that caption to the exact same embedding call naive_rag Lesson 2 used for a text chunk. That's the entire payoff of choosing captioning-then-embed back in Lesson 1: once an image becomes a caption, it is, as far as embedding is concerned, indistinguishable from any other piece of text in this series.

The code, piece by piece

caption = caption_image(IMAGE_PATH)
vector = embed_texts([caption])[0]

embed_texts here is naive_rag Lesson 2's embed_content call, copied verbatim, same model (models/gemini-embedding-001), same output_dimensionality (768). No parameter changes, no "image mode", because by the time this function runs, there is no image left, only the caption string caption_image produced.

print(f"Vector length: {len(vector)}")

768 numbers, exactly the same shape a text chunk's embedding has. This is worth confirming directly rather than taking on faith: it's what makes Lesson 6's mixed store possible, a text chunk's vector and an image caption's vector can sit in the same list and be compared with the same cosine similarity function, because nothing about their shape tells them apart.

Checkpoint

  • Once an image is captioned, its embedding uses the identical call, model, and dimensionality as any text chunk in this series.
  • A caption's vector and a text chunk's vector are the same shape, there is nothing in the vector itself that marks it as "came from an image." Lesson 11 adds that information back as metadata, on purpose, once it turns out to matter for citing sources correctly.

If anything here still feels unclear, ask before moving to Lesson 6.