Uploaded Documents (PDF/DOCX) as a data source: RAG with images, links, and ontology

When important knowledge is trapped in PDFs

If your knowledge already lives in PDF / DOC / DOCX (manuals, policies, catalogs, training), it often behaves like an “archive”: everything is there, but getting the exact answer fast can be surprisingly hard. You want the agent to reply confidently — with proof, with a link to the source, and when it helps, with the exact screenshot/diagram right inside the chat.

That’s what the Uploaded Documents data source in Lite.panteo.ai does. You upload files in the dashboard, and the platform builds a RAG index: it extracts text, splits it into chunks, attaches metadata (document link + page), extracts illustrations, and (optionally) enables ontology-aware search (sections, headings, definitions, tables, lists). The result feels very different: the agent doesn’t say “somewhere in the PDF” — it answers with the fragment + the image + the page link.

Common friction points this solves

• The documents exist, but finding the exact place where something is stated takes time.
• Support and sales end up answering similar questions over and over from the same materials.
• “Where is this button?” questions often need a visual — and the conversation turns into sharing screenshots.
• Documents get updated, while a Q&A base (if you have one) doesn’t always stay in sync.

With Uploaded Documents, your files stay the single source of truth — and the agent does the retrieval work, including links and visuals.

Use cases: where this shines

Help center and product documentation

Upload PDF manuals and technical docs. The agent retrieves the relevant passage by meaning and can surface a diagram or UI screenshot from the PDF plus a link to the exact page — fewer tickets, faster resolution, higher trust.

Policies, compliance, and regulations

When structure matters more than keywords, switch to Level 3 (ontology). The agent navigates sections and subpoints, tracks definitions, and handles tables/lists more reliably in long documents.

Training and onboarding

Training PDFs often include tables and diagrams. With document RAG, the agent answers by meaning and can return the right image/table in the reply with a direct source link.

Proposals and catalogs for sales

Turn proposals and catalogs into a searchable asset: the agent finds specs, terms, and configuration details, can show a product image or a catalog page, and links the original document for review.

Why answers feel “verified”: 3 key capabilities

1) Images from PDFs — inside the reply

The system extracts images from PDF (and, when supported, from DOC/DOCX) and attaches them to the right chunks. That means search results include both text and linked illustrations, so the agent can insert the right diagram/screenshot/photo into the chat response.

2) Metadata and links: document + page

Each chunk stores source_url (public file link) and source_page (page number). So the agent can answer with “see p. 5” and provide a link — users can verify instantly, without hunting for the PDF.

3) Ontology (L3): when structure is the answer

Just like the Google Docs source and the web parsing data source, uploaded documents support ontology-level indexing: sections, headings, definitions, tables, and lists. This is what makes long policies and regulations searchable “by meaning and structure”, not only by matching words.

How to set it up in the UI

Setup in ~10 minutes: from file to answer

Step 1: Create an Uploaded Documents source

1. Go to Dashboard → Data Sources → Add Source.
2. Choose Uploaded Documents.
3. Enter a name and description (e.g. “Manuals and policies in PDF”).

Step 2: Upload PDF/DOC/DOCX

4. On the Main (or Files) tab click Upload documents or drag and drop files into the upload area.
5. Supported formats: PDF, DOC, DOCX. You can upload multiple files at once (max file size 60 MB).
6. After upload, the list shows document name, public URL, and size.

Step 3: Enable images and source links

7. Check Include illustrations — images from documents will be extracted and attached to chunks; the agent can show them in replies.
8. Check Include source links and optionally Append source link to chunk text — replies can include links to the document and page.

Step 4: Choose search depth (L1/L2/L3)

9. Level 1 — Lexical search: keyword search, fast, minimal cache.
10. Level 2 — Semantic search (recommended): embeddings, meaning-based search.
11. Level 3 — Ontology search: uses document structure (sections, definitions, tables, lists) for complex policies and long texts.

In Advanced you can set chunk size, chunk overlap, and search agent refinement iterations.

Step 5: Run indexing and review chunks

12. Click Refresh / Re-index. The system parses documents, creates chunks, extracts illustrations, and builds the selected index level (L1/L2/L3).
13. In the Chunks tab you can review fragments, metadata (source_url, source_page), attached illustrations, and ontology.

Step 6: Attach the source to your agent

14. Open Dashboard → Agents and select the agent.
15. In Form fields (or Data sources / Knowledge) link the Uploaded Documents source you created.
16. Save the agent.

Step 7: Test it in the chat

17. Open the chat widget or Telegram and ask questions that should be answered from the uploaded documents.
18. Confirm that answers are accurate and that the agent uses illustrations and document links when appropriate.

Summary of setup parameters

Parameter	What it does
Files	PDF, DOC, DOCX up to 60 MB; upload via button or drag-and-drop.
Include illustrations	Extract images from documents and show them in agent replies.
Include source links	Add document and page links to search results.
Indexing level	L1 — lexical, L2 — embeddings (recommended), L3 — ontology.
Chunk size	Amount of text per fragment (e.g. 400 tokens).
Chunk overlap	Overlap between adjacent chunks for continuity.
Refresh	Re-run parsing and indexing after changing files or settings.

Try it with one PDF (you’ll feel the difference fast)

If you want the agent to answer from your documents (not generic web knowledge), start small: upload one PDF, enable illustrations, pick an indexing level (usually L2, and for policies L3), and ask a couple of real questions in the widget. You’ll immediately see what changes: sharper retrieval, images in the reply, and links to exact pages that make answers feel trustworthy.

Register in Lite.panteo.ai and connect Uploaded Documents — it’s one of the fastest ways to turn PDF/DOCX into a working RAG knowledge base for your agent.

Related articles

• Combine with API-powered data sources when you need live data alongside documents.
• Use Google Docs as a data source for cloud documents by link; uploaded documents for local PDF/DOC/DOCX.
• The Web parsing data source is for documentation on a website with image extraction; uploaded documents for files.
• Keep consistent answers in the Q&A knowledge base; use uploaded documents for full text and illustrations.
• Send leads from the chat to your CRM via CRM integrations.
• Lead management: Lead management with full dialog history.