How to connect an n8n agent to PDF documents without vector databases

n8n architecture comparison: vector pipeline vs direct ingest with Claix

Why vector databases complicate n8n flows

Integrating documents into n8n agents through vector stores creates common operational friction in production:

1. Unstable tool invocation

When a vector store is wired as an agent tool in n8n, the LLM often skips querying the database or answers from internal knowledge if the tool description is imperfect. That introduces non-determinism in flows that must process documents every time.

2. Broken numeric tables and balances

Traditional text-splitting nodes (Recursive Character Text Splitter) do not understand table layout. Cutting a PDF every 1,000 characters separates an invoice row from its column headers, so the agent extracts wrong or incomplete amounts.

3. Configuration overhead and hidden costs

A RAG flow in n8n requires managing embedding model credentials, configuring vector indexes, controlling vector dimensions, and paying for vector-database read/write operations on every automation run.

Performance and operating-cost comparison per flow

Practical HTTP Request node setup in n8n

You do not need complex community nodes or Python scripts. Configure a single standard HTTP Request node:

┌─────────────────────────┐     Binary Data     ┌─────────────────────────┐     Structured JSON     ┌─────────────────────────┐
│     Trigger Node        │────────────────────►│  HTTP Request (Claix)   │────────────────────────►│     Next action         │
│ (Email / Form / Drive)  │                     │ POST /api/pdf-json      │                         │ (PostgreSQL, CRM, etc.) │
└─────────────────────────┘                     └─────────────────────────┘                         └─────────────────────────┘

n8n configuration parameters

• Method: POST
• URL: https://claix.dev/api/pdf-json
• Authentication: Generic Credential Type → Header Auth (or manual x-api-key header).
• Send Body: True
• Body Content Type: Form-Data (multipart/form-data)
• Body Parameters: schema_id (String) — the UUID of the schema previously created in Claix.
• Body Parameters: file (Binary Data) — the binary property name that holds the PDF (for example, data).
• Body Parameters: space_id (Optional) — the space UUID if the document belongs to a shared document set.

Data structure: input vs output on the n8n canvas

1. Output payload from the Claix HTTP Request node

After the file is processed, the node emits structured JSON ready for any downstream node (database, Google Sheets, Slack, or CRM) without writing regular expressions:

{
  "success": true,
  "schema_utilizado": "Supplier Invoice",
  "total_registros": 1,
  "document_id": "d4a1e9d2-8b1c-4f3e-9a02-8b1e9f3c7a4b",
  "data": [
    {
      "numero_factura": "F-2026-089",
      "proveedor": "Logística Global S.L.",
      "fecha_emision": "2026-03-12",
      "base_imponible": 4500.00,
      "iva_importe": 945.00,
      "total_factura": 5445.00,
      "lineas": [
        {
          "concepto": "Container transport Valencia-Madrid",
          "cantidad": 3,
          "precio_unitario": 1500.00
        }
      ]
    }
  ]
}

2. Follow-up queries with a later HTTP node

If the n8n agent needs another question about the PDF (for example: "Is there a late-delivery penalty?"), it does not re-upload the binary. Add a second HTTP Request node:

• URL: https://claix.dev/document-context/{{ $json.document_id }}
• Body (JSON): { "questions": ["Is there a late-delivery penalty, and how much is it?"] }

{
  "questions": [
    "Is there a late-delivery penalty, and how much is it?"
  ]
}

Common use cases in n8n