What you can ask

Once you've uploaded a document under one of the canonical financial doc types (trust deed, building contract, will, partnership agreement, super PDS, lease, mortgage variation, QS depreciation report, etc.), Frank can answer questions about its content.

Examples:

• "What does my SMSF trust deed say about distributing capital to a beneficiary under 18?"
• "What's the milestone payment schedule in my building contract?"
• "When does my lease end and what's the bond amount?"
• "Does my partnership agreement allow a unilateral exit?"
• "What's the FBT exemption status of the vehicle in my novated lease quote?"

Frank returns the most relevant excerpts, with the source document name and a similarity score. You can then click through to view the full document for context.

How it works

After OCR completes for a canonical-type document:

1. Frank chunks the extracted text into ~1000-character windows with 200-character overlap (so meaning isn't sliced at boundaries).
2. Each chunk is embedded with text-embedding-3-small (the same model that powers our AU tax knowledge base).
3. Embeddings are stored per-user in a vector index with HNSW for fast cosine-similarity search.

When you ask Frank a question that involves your documents, the search_user_documents tool embeds the query and pulls the top matches.

Type-aware filtering

You can scope a search to a single document type:

"Frank, in my trust deeds only — what's the rule about lump-sum distributions to children?"

Frank uses the type filter to keep results focused.

Not embedded by default

Frank only embeds canonical financial doc types automatically. Receipts, transactions, bank statements stay grep-searchable but aren't vector-embedded (cost control — these are usually short and substring search is enough).

If you want to force-embed an existing document, hit the embed button in the document detail panel, or run Reindex all my documents in your Documents settings to backfill everything that qualifies.

Privacy

Embeddings live in your per-user table with the same RLS as your raw documents. Practice members can read embeddings only for documents you've toggled practice_visible (or uploaded as tax_document).