Skip to main content

Indexing

Before Assistant can reason over a file, Nomic indexes it: the file is parsed into structured text and images, then embedded for semantic search. Indexing is a one-time cost per file version — after that, every question Assistant answers about the file uses only foundation-model tokens.

How indexing works​

Indexing runs two models under the hood:

  1. Nomic Parse — Nomic's domain-specific parsing model turns a document into structured text, tables, and figure extractions. For drawings and PDFs, Parse handles sheet segmentation, OCR, and layout understanding.
  2. Nomic Embed — Parsed content is chunked and embedded so it can be retrieved by semantic search.

Once a file is indexed, its parsed content and embeddings are stored and reused across every session, workflow, and search. Re-indexing only happens when a new version of the file is uploaded or synced. You can see each file's indexing status — Indexed, Indexing, Failed, or not yet indexed — in the Status column of the Files view. See Files for the full UX.

Supported file formats​

  • PDF — including multi-sheet drawing sets, specifications, and reports. PDFs auto-index after upload or sync.
  • Microsoft Word — legacy .doc files auto-index. Open XML Word files (.docx, .docm, .dotx, .dotm) can be indexed with Index Document or Index Folder, but do not auto-index on upload.
  • Microsoft Excel (.xls, .xlsx, .xlsm, .xltx, .xltm) — index with Index Document when Assistant needs to search the spreadsheet.
  • Microsoft PowerPoint — .ppt, .pptm, .potx, and .potm files auto-index. .pptx files can be indexed with Index Document or Index Folder, but do not auto-index on upload.
  • BIM models (.ifc, .ifczip) — prepared on demand for BIM viewing and Assistant queries
  • AutoCAD drawings (.dwg, .dxf) — viewable in the CAD viewer, but not indexed for semantic search
  • Google Docs, Sheets, and Slides and images — stored and viewable, but not indexed for semantic search

Syncing a file from an integration does not index it. SharePoint, Autodesk Forma, Bentley, and Fieldwire syncs bring in files, folders, and permissions; indexing is a separate step you trigger. Egnyte syncs are the exception — an admin can enable auto-indexing when configuring the sync.

Index on demand with Index Document in a file's action menu, or Index Folder on a folder — for yourself only, or for everyone in the organization. A folder request keeps picking up files that arrive later, so running Index Folder once on a synced folder covers ongoing sync. Folder indexing skips spreadsheets, BIM models, and CAD drawings.

BIM model preparation​

BIM models do not follow the normal Parse and Embed indexing flow. Nomic prepares IFC-derived model artifacts when you open a model, attach it to Assistant or workflow context, or ask Assistant to query it, then caches those artifacts for the file version. Because BIM models are not split into document pages, they do not show the same document indexing status or page-based indexing estimate as PDFs and Office files.

Pricing​

Indexing cost is driven by pages parsed, not the number of files:

StepUnitPrice
Nomic ParsePer page$0.012 / page
Nomic EmbedPer fileIncluded in the parse cost for typical documents

Reference costs:

  • A 10-page spec costs about $0.12 to index.
  • A 100-sheet drawing set costs about $1.20 to index.
  • A 1,000-page project archive costs about $12 to index.

Each parse event is recorded as a spend event against your AI Usage pool with the page count and cost. Admins can review parse events across the organization in Admin → Usage or Admin → Analytics → Parsing; individual users can review their own parse events in Settings → Usage.

The per-page price absorbs layout segmentation, OCR for scanned or mixed-content PDFs, table and figure extraction, drawing sheet decomposition, and embedding for both text chunks and image regions. There is no separate per-token charge.

Indexing vs. not indexing​

You can attach a non-indexed file directly to a single Assistant session, and Assistant will read it for that conversation. The tradeoffs:

IndexedNot indexed (attached only)
Retrievable across sessionsYes — available in every thread and workflow in the projectNo — only visible to the session it is attached to
Searchable via semantic searchYesNo
Scalable to large setsYes — Assistant pulls only relevant chunksNo — the whole file goes into the model's context
Cost per questionLow — only pays for the tokens Assistant actually usesHigh — every question re-reads the full file, paying foundation-model input tokens each time
One-time costYes — $0.012 / page, then reused foreverNone upfront, but every turn pays for the full file

The break-even point is low. For a 50-page document, indexing costs $0.60 once. Reading that same file inline on Sonnet 4.6 costs roughly $0.05 – $0.15 per question in input tokens alone. After a handful of questions, indexing has paid for itself — and the file is then available to every future session and workflow.

Rule of thumb: index anything you expect to reference more than once, or anything that will live in a project the team will return to.

Managing indexing costs​

  • Index selectively. You don't need to index every file in a SharePoint site or Autodesk Forma project. Use integration configuration to scope which folders sync into Nomic.
  • Use personal vs. organization indexing for one-off exploration to avoid promoting files to canonical project context.
  • Avoid re-uploading. Re-indexing is tied to file versions — uploading the same file twice indexes twice.
  • Watch large binary PDFs. Sort the Files view by page count to find the most expensive files to index.
  • Audit indexing spend. Admins can filter Admin → Usage to parse rows or open Admin → Analytics → Parsing to see parse spend by user and individual parse events. Individual users can open Settings → Usage to see their own parse events.
  • Scope Assistant context. Even after indexing, narrowing context with projects and folders keeps per-question costs in check.

Where to next​