After reading "25 Years of Eggs", where a careful multi-stage pipeline of SAM3, PaddleOCR, and Claude handled 11,345 receipts, I was curious how a smaller vision-enabled model would fare going directly from image to structured data, with no preprocessing at all.

I put together a quick example babashka script for receipt extraction using clj-llm:

(def ai (openrouter/backend
         {:defaults {:model "google/gemini-3-flash-preview"}}))

(def ReceiptSchema
  [:map
   [:currency :string]
   [:items [:vector
            [:map
             [:name {:description "expand all abbreviations, convert to english"} :string]
             [:category {:description "eg. dairy, produce, meat etc"} :string]
             [:value [:map
                      [:unit {:description "eg. kg, per-unit etc"} :string]
                      [:price :double]]]
             [:quantity-of-value :double]
             [:total-price :double]]]]
   [:total :double]
   [:date :string]
   [:store :string]])

(def system-prompt "Extract structured data from the receipt image and return it in the specified schema format.")

(llm/generate ai
 {:schema ReceiptSchema
  :system-prompt system-prompt}
 [(content/image (first *command-line-args*)) {:max-width 512}]) 

I was able to get clean, structured output from a variety of test receipts. Obviously this is just a simple muck-around and getting it up to the level the original author had would take far more structured test input and iteration, but I think it's a promising start.

Even more promising is the difference in cost - using the small gemini model, a single receipt costs only around $0.002.

{:cost-details
  {:upstream-inference-cost 0.00214,
   :upstream-inference-prompt-cost 6.01E-4,
   :upstream-inference-completions-cost 0.001539},
  :finish-reason "tool_calls",
  :completion-tokens 513,
  :prompt-tokens 1202,
  :prompt-tokens-details
  {:cached-tokens 0, :cache-write-tokens 0, :audio-tokens 0, :video-tokens 0},
  :cost 0.00214,
  :is-byok false,
  :total-tokens 1715,
  :completion-tokens-details
  {:reasoning-tokens 0, :image-tokens 0, :audio-tokens 0},
  :model "google/gemini-3-flash-preview"}

Making the total for the original article's 11,345 receipts come out to $22.69. A fair reduction from his original $1,591.

We get some other nice benefits too:

• Translation is an LLM's bread and butter. We can normalize to English easily.
• Abbreviated terms can be expanded by the LLM without too much fuss.

The script would benefit from testing with a larger sampling of receipts, and using an enum to keep the categories regular across different stores. Notice how the target splitting of GROCERIES and HEALTH-BEAUTY-COSMETICS has a large influence on the way the categories are output. Adjusting the schema with a better description can help, but this is better achieved with a sampling of your common receipts and use of an enum to reign-in the required categoires you need.