Large Markdown documents

Zelares can store Markdown documents that are much larger than the amount of text a single AI operation should inspect. Document storage, search indexing, and AI context therefore use separate limits and should not be treated as one quota.

Document size limits

These are document-save limits. They are separate from Workspace storage capacity and from the AI context limit described below. Internal storage thresholds are implementation details and do not replace your plan's document-size entitlement.

How large Markdown is stored

Zelares may keep smaller Markdown directly with the document record and move larger Markdown to external object storage. That placement is transparent: you continue opening and editing the same document, and the internal storage threshold is not a separate plan limit.

Existing documents do not require a bulk migration. Older inline documents remain valid and can transition through normal saves and size changes when needed.

Chunks, search, and indexing

Large Markdown can be represented as a complete set of smaller derived chunks for retrieval and indexing. Chunk generation covers the current document; it does not stop after the first 12 MiB. Relevant chunks may therefore be selected from the beginning, middle, or end of a large document.

• Chunks and retrieval indexes are derived data, not the authoritative copy of your Markdown.
• Indexing can temporarily be pending or rebuilding after a change.
• A derived index can be rebuilt without changing the saved Markdown document.
• Older or incomplete derived generations are not silently treated as the current document.

AI context is limited per operation

This section documents the server-side context boundary that any model or tool integration must preserve. Zelares does not currently expose a general AI assistant or model-completion runtime.

One logical AI operation can receive at most 12 MiB total of authoritative Markdown text selected from the current document chunks. The 12 MiB ceiling is aggregate across all selected chunks and all retrieval passes in that operation.

up to 100 MB Markdown document
        ↓
complete derived chunk coverage
        ↓
relevant chunks selected from anywhere
        ↓
≤ 12 MiB total Markdown context for one AI operation

• 12 MiB is not a per-chunk allowance.
• It is not limited to the first 12 MiB of the document.
• Multiple retrieval passes in one logical operation share the same 12 MiB ceiling.
• A provider or model may require a smaller effective context limit.
• A later independent AI operation may retrieve a different region of the same document.

If selected context does not cover the complete current document, it is partial context. AI-facing behavior must not claim the entire document was inspected when omitted regions were not supplied.

When indexing is pending

If retrieval reports that chunks or embeddings are still indexing, the saved Markdown remains authoritative. Retry the operation after indexing has had time to rebuild. A derived-index failure should not require you to recreate or split the document solely to preserve its saved content.

Common questions

Does AI automatically read an entire 100 MB document?

No. Retrieval selects relevant current chunks, and one logical operation receives at most 12 MiB total Markdown context.

Is the 12 MiB limit applied to every chunk?

No. It is one aggregate ceiling shared by all selected chunks and retrieval passes in the operation.

Does chunking stop once 12 MiB has been indexed?

No. The derived chunk corpus can cover the complete current document so later operations can retrieve different regions.

Related pages

For Markdown syntax and editing, see Markdown Basics. For Workspace capacity and billing guidance, see Usage and billing.

The internal AI context constant is 12 MiB per logical operation.