Skip to main content

Optimization with LiteParse & Prompt Caching

To minimize API operating costs and accelerate bulk document parsing, PDFHUB integrates two core optimization technologies: the local open-source document parser LiteParse and API Prompt Caching.

1. LiteParse: Local Document Parsing​

While LlamaParse Cloud offers advanced parsing capabilities, PDFHUB also supports LiteParse running directly on your local CPU (no GPU required) for situations requiring strict data privacy or zero-cost execution:

  • Model-Free Parsing: LiteParse uses spatial geometry algorithms to identify lines, columns, and text blocks directly from the vector code of PDF files.
  • Local OCR Support: For scanned pages, LiteParse integrates lightweight OCR tools (such as PaddleOCR or Tesseract) to recognize text without sending data online.
  • Spatial Layout Preservation: Extracts text along with bounding coordinates (x,y,w,h)(x, y, w, h) for each element, keeping the original spatial structure.
  • CLI Support: Run quick parsing tasks via the CLI on your server: 
    liteparse parse --file report.pdf --format markdown --output result.md

2. Prompt Caching: Minimizing API Overhead​

When users perform multiple RAG queries or data extractions on the same large PDF document (e.g., querying details across a 150-page corporate financial report), sending the entire document context repeatedly to cloud LLMs (such as Anthropic Claude or DeepSeek) is expensive. PDFHUB addresses this using Prompt Caching:

Query 1 ──► Send Full Context (Prompt + 150-page PDF) ──► Server Processes & Caches (100% cost)
Query 2 ──► Send New Question + Cache Token ID ────────► Server Reuses Cached State (Saves ~80% cost)

How It Works​

  1. Prefix Matching: The system splits outgoing prompts. The heavy PDF document context is positioned at the beginning of the prompt sequence (System Prompt / Context block).
  2. State Caching: On the first query, the model provider (e.g., Anthropic/DeepSeek) compiles and stores this context state in their server-side cache.
  3. Cache Hits: In subsequent queries within the same chat session, the system only sends the new question and a cache token identifier. The model reuses the cached context, processing the response immediately.

Economic Benefits​

  • Reduced Input Cost: Saves 50% to 80% of input token costs for documents exceeding 10,000 words.
  • Lower Latency (Time-to-first-token): The time to begin generating the response is 3 to 5 times faster since the model does not need to re-read the entire document.