Products

How Bruq AI Indexes PDFs on Android: OCR, Citations, and BYOK

Published August 9, 2026 · Influrion Editorial Team

Bruq AI (formerly BrainVault) is Influrion Solutions’ Android knowledge vault. This post walks through the pipeline in product language — what happens from import to a cited answer — without claiming magic or “fully offline chat.”

The user journey

  1. Unlock with the phone’s PIN, pattern, fingerprint, or face.
  2. Build the vault — Add more, pick files, or let Bruq AI find PDFs/images; optional auto-recognition watches folders like Downloads, Documents, and Pictures.
  3. Index on device — extract → chunk → embed until items show Ready.
  4. Ask or search — Chat with Verified Sources, or hybrid / semantic / keyword search.
  5. Verify — open citation text and the original document.
  6. Tune — API keys, ingest confidence, watched folders, System Integrity grounding in Settings.

Pipeline (what runs where)

Phone storage / picker
        ↓
   Import + triage
        ↓
 OCR / PDF extract  →  chunk  →  on-device embed
        ↓
   Local vault index (search + history)
        ↓
 Ask / Search  →  retrieve snippets  →  (optional) LLM answer + citations
  • Local: PDF text extraction, image OCR, chunking, embeddings, SQLite index (FTS5 + vectors), vault unlock, API key storage.
  • Only when you ask with a provider configured: question + retrieved snippets → Groq (primary) or OpenRouter (fallback).

Hybrid search and citations

Keyword search alone misses paraphrase. Vectors alone can miss exact IDs and invoice numbers. Bruq AI fuses both for better recall, then grounds chat answers in retrieved chunks. Citations are meant to be inspectable — a trust feature, not decoration.

Grounding rule (exposed as System Integrity): answer from provided context; otherwise say you do not know.

BYOK: your keys, your provider

Influrion does not ship hosted LLM keys inside the binary. You paste Groq and/or OpenRouter keys in Settings. Temperature controls response style. That keeps cost and provider choice with the user — and keeps the vault from becoming Influrion’s cloud.

Smart photo filtering

Optional on-device labeling can skip selfies, family photos, landscapes, and food images during auto-recognition. Filtered items appear as Unsupported, not as vault errors.

Stack (brief)

Expo SDK 56, React Native, on-device SQLite, ML Kit OCR, local embeddings, Groq / OpenRouter for answers. Native modules — not Expo Go as the consumer product.

Get the app

Keep reading

Need a similar on-device or BYOK architecture for your product? Contact Influrion Solutions.