Skip to content

Drafting quick wins

complydoc assist reads a report an audit already wrote and asks a hosted chat model what to do about it: where readiness is being lost, and which loader read the folder best.

complydoc audit ./documents --out build --name current
complydoc assist --report build/current.json

- Turn on OCR for the three scanned invoices
  They have no text layer, so nothing was read from them and they were not scanned
  for identifiers.

Every other report in complydoc is computed on the machine it runs on, this one included: quick_wins[] in the JSON is written by an audit without any model. assist is a second opinion on top of that, and it is the part that leaves.

What it sends

The report, as JSON, to the model you name. That is:

Sent Held back
Findings, with identifiers masked The text read off each page
Readiness signals and their reasons The page pictures, if the run embedded any
Routing, loaders, extractions, timings Every priced model against every document
Cost totals, limitations, run metadata
Document paths and file names

Masked identifiers and the paths of your documents are in what goes. A path can name a client. If that matters, audit a copied folder with neutral names, or do not use this command.

Holding back the page text and the page pictures is not only about privacy. The ten-page sample report is 509,000 tokens if sent whole with pictures, and 52,000 once those parts are dropped. What remains grows with the folder, at roughly 5,000 tokens a document, so a large folder needs --sample on the audit that wrote the report, or a model with room for it. The command says so before it sends anything large.

Install and run

uv tool install --force "complydoc[assistant]"
export OPENAI_API_KEY=...        # whatever the model you name expects
complydoc assist --report build/current.json --model gpt-5.6-luna

The extra brings LangChain, which resolves --model through init_chat_model, so any provider LangChain supports works if its package and its key are present.

Before the call, the command names the model it is about to send to. After it, it lists the hosts that were reached.

From Python

import complydoc as cd
from complydoc.integrations.assistant import quick_wins_call

report = cd.load_report("build/current.json")
message = quick_wins_call(report, allow_network=True)
for win in message.quick_wins:
    print(win.quick_win, "-", win.justification)

allow_network=True is required and defaulted nowhere. Without it the call raises ValueError rather than quietly sending the report. connections_made() returns the hosts reached.

What it does not do

  • It does not run as part of audit, check or a policy gate.
  • It does not write the drafted wins into the report. What it prints is the model's answer, not a finding complydoc stands behind.
  • It does not change run.content_sent_to in the report it read. That field records what the audit itself sent, which is nothing.