Skip to content
Retrieval

Visual document retrieval (late interaction)

Search scanned forms, slides, drawings and table-heavy PDFs as images, with no OCR pipeline to maintain.

Use it when

Corpora where layout carries meaning: engineering drawings, slide decks, scanned contracts, invoices and financial tables that OCR-based pipelines flatten and lose.

Structure

The parts, top to bottom

hover a part to see its job

across every level

Hover or tap any part to see what it does. The light shows the order a request moves through.

Flow

What happens, in order

  1. 1Rasterise each page
  2. 2Encode into patch embeddings (ColPali / ColQwen)
  3. 3Store multi-vectors with MaxSim support
  4. 4Query encoded token by token
  5. 5Top pages passed to a vision model
  6. 6Answer with page citations
Tools

What we typically build it with

ColPali / ColQwenVespa / QdrantGemini / GPT / Claude visionLanceDBRAGAS

Trade-offs

Higher storage per page (about a thousand patch vectors) and heavier queries, but ColPali scored nDCG@5 of 81.3 on the ViDoRe benchmark versus roughly 65–75 for OCR pipelines. Worth it when documents are visual; unnecessary for clean text.

What we solve

Problems this architecture solves

Generic problem statements with the flow and the outcomes the industry has documented.

Document intelligence pipeline

Documents keyed in by hand

Invoices, KYC files, claims and contracts arrive as PDFs and photos. People retype them, errors slip through, and the backlog grows every month-end.

How the system works

  1. Ingest PDF / image / email
  2. Classify document type
  3. Extract to a schema
  4. Validate & score confidence
  5. Auto-post or route to review
Multimodal LLMStructured outputsReducto / LlamaParseReview queueERP / LOS integration

Outcome: Documented deployments classify documents in under a second, turn day-long claims backlogs into minutes, and reach 85% no-touch invoice processing within six months with a seven-month payback. DXC with Claude; Vic.ai

Architecture
Visual document retrieval

Answers locked in scans, slides and drawings

Engineering drawings, slide decks, scanned contracts and table-heavy PDFs lose their meaning when flattened to text, so search returns nothing useful.

How the system works

  1. Rasterise pages
  2. Patch embeddings (ColPali)
  3. Multi-vector index
  4. Retrieve pages, not chunks
  5. Vision model answers with citations
ColPali / ColQwenVespa / QdrantGemini / GPT / Claude visionLanceDBGolden set

Outcome: Late-interaction retrieval over page images scored 81.3 nDCG@5 on the ViDoRe benchmark against roughly 65–75 for OCR pipelines, with no OCR stack to maintain. ColPali (ICLR 2025)

Architecture
Related

Other retrieval patterns

Retrieval

Hybrid RAG with reranking and citations

Keyword plus vector search, fused and reranked, answered with sources.

Use it when: Questions over a body of documents, tickets or records where exact terms matter as much as meaning, and where every answer must be checkable and permission-aware.

Flow

  1. 1Ingest & chunk by structure
  2. 2Enrich with context & metadata
  3. 3Embed + keyword index
  4. 4Query: BM25 + vectors, fuse (RRF)
  5. 5Rerank top candidates
Retrieval

Agentic and graph-based retrieval

The agent decides how to search; a knowledge graph answers the global questions.

Use it when: Multi-hop questions, comparisons across many documents, or questions about themes across a whole corpus where chunk retrieval alone returns fragments.

Flow

  1. 1Question analysed
  2. 2Agent chooses strategy: search, graph, SQL, or ask
  3. 3Runs several retrievals
  4. 4Verifies & consolidates evidence
  5. 5Answers with a reasoning trail

Let's build intelligent systems that drive growth

Tachyon is the engineering partner for teams that need AI in production, not in a deck. Start with a free 60-minute discovery call.

Contact

Tell us the problem, we will map it to the architecture

Every great partnership begins with a conversation. Whether you are exploring possibilities or ready to scale, tell us what you are actually trying to build.

Prefer to talk?

Pick a 60-minute slot. No pitch, just an engineer with honest answers.

Book a call

NDA available on request. We reply within one business day.