Use DeepDoc

Use DeepDoc

What it costs

Which commands call an LLM, which are free, and how to measure your own spend.


Most tools in this space are vague about cost. Here is the whole picture.

Free forever

These commands never call an LLM. Run them as often as you like.

Command
deepdoc serveRebuilds the site from your saved plan
deepdoc deployBuilds static output
deepdoc statusReads local state
deepdoc configReads and writes YAML
deepdoc performanceReads a local log
deepdoc cleanDeletes files
deepdoc initWrites a config file

Every theming change is in this list. Colours, presets, fonts, logo, sidebar order, renaming pages, hiding pages, adding your own — all rebuilt from .deepdoc/plan.json, none of it re-reads your code.

What costs money

Two commands.

deepdoc generate

The full run. Roughly:

Scan — no LLM. Parses your code locally.
Plan — a small number of calls to name and group the clusters found in your call graph.
Generate — approximately one call per page, in batches of 10 with up to 6 in flight.
Consistency pass — one further pass adding cross-page links. Disable with consistency_pass: false.

So the dominant term is pages, not repository size directly. A repo that plans to 40 pages costs roughly 40 generation calls plus planning overhead.

A page that fails validation is retried, up to MAX_RETRIES of 3. Most pages never retry; a page that does costs 2–3× that page.

deepdoc update

The same machinery, but only for pages affected by your diff. Change one module and you pay for the handful of pages that describe it, not all 40.

This is why update is the normal command after your first generate.

Measuring your own

Do not trust an estimate — including one from these docs. Measure it:

deepdoc performance

That reads .deepdoc/performance/runs.jsonl, which records per-run timings, token counts and retries. The token counts there are the real numbers for your repository and your model, which is the only figure that matters.

Spending less

Use `update`, not `generate`

After the first run, update regenerates only what your diff touched. This is the single biggest lever.

Use a cheaper model

Set llm.model to a smaller model. Documentation is mostly summarisation, which smaller models handle well. See Providers.

Run a local model

Ollama costs nothing per token. Slower, and quality varies by model, but the marginal cost is zero.

Narrow the scope

exclude in .deepdoc.yaml keeps generated tests, vendored code and build output out of the plan. Fewer files means fewer pages means fewer calls. Unlike theming, this changes what gets documented, so set it before generating.

Turn off the consistency pass

consistency_pass: false skips the cross-linking pass. You lose the “See also” links between related pages.

Edit this page on GitHub