Table extraction from PDFs is central to our pipeline, but since ~every tool is imperfect, we end up going to an LLM to batch-correct all of the rows in our extracted tables. The current prompt looks like the below, where each column header is a field on

CorrectedData

. The big issue is that this takes a long time! Any suggestions for how I might make this process faster without sacrificing accuracy?

class CorrectedData {
  @@dynamic
}

class CorrectedRowsData {
  rows_corrected int @description("How many rows were corrected/modified from the original Camelot output")
  correction_notes string[] @description("VERY BRIEF notes about corrections made (e.g., 'merged cells in rows 5-7', 'removed phantom column 3'). MAX 200 chars total.")
  corrected_data_rows CorrectedData[] @description("ALL data rows corrected to match visual. Each inner array has exactly len(column_headers) cells in correct order. Headers already removed. Each row MUST have a non-empty value in the 'code' column (door mark/number or unit type code) to be included.")
}
function CorrectTableRows(
  camelot_rows: string[],
  broad_schedule_type: "unit" | "door" | "window",
  schedule_image: image
) -> CorrectedRowsData {
  client Gpt4_1
  prompt #"
    You are correcting table rows from a {{ broad_schedule_type }} schedule.

How large are your requests right now? Like how much data are you feeding into the request

We show the whole thing to the user after it's done becuase it's all in an async pipeline. We generally feed in ~25 rows of data, with anywhere from 5-20 columns.

How big is each row?

Pdfplumber is pretty good at table extraction

I'm looking for more baml-y tricks. One example I got from @Vaibhav (BAML 🐑) is-- add "no quotes on keys" in the json output, which saves time. Anything like that?

Most of it really goes to just simplifying your schema