{
  "id": 15350302,
  "title": "Why parsing real estate contracts with LLMs is ten times harder than tutorials claim",
  "url": "https://urgent.news/2026/10/18/why-parsing-real-estate-contracts-with-llms-is-ten-times-harder-than",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-18T04:52:15.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/novasiaestate/why-parsing-real-estate-contracts-with-llms-is-ten-times-harder-than-tutorials-claim-4pak"
  },
  "original_language": "en",
  "account": "When our team began constructing the data pipeline for NovAsia, a Southeast Asia property catalog and market tool, the initial plan seemed straightforward on paper: acquire property PDFs and developer contracts, feed them into a cutting-edge vision LLM, and output structured JSON into PostgreSQL. However, after just two weeks, the pipeline became entirely dysfunctional. This is why the standard tutorial guidance fails in real-world production, and the adjustments we made to achieve dependable data extraction from scanned documents. What most tutorials neglect to mention Most examples in LLM document parsing tutorials utilize pristine, digital PDFs with neat typography. In emerging markets like Cambodia and Thailand, the situation is far more complex: Phone scans dominate, with half the documents being multi-generation scans or photographs taken with a smartphone under suboptimal lighting conditions. Dual-language layouts also present challenges, as Khmer or Thai text is frequently intermingled with English, often occupying adjacent table columns where line baselines do not align. Physical stamps, such as red rubber stamps, signatures, and notary seals placed directly over crucial numerical data (payment milestone dates, sinking fund fees, transfer taxes), add another layer of complexity. The 40-page prompt dilemma: Loading a 35-page PDF into a long-context model and requesting all fees and payment dates leads to 80% accurate extraction, but the remaining 20% are confidently fabricated by the model. Three essential modifications that resolved the issue Rather than attempting to find a single model that can handle everything, we decomposed the pipeline into more manageable, straightforward steps: Aggressive image enhancement prior to text extraction Before any model processes the page, we employ basic OpenCV techniques: deskewing, binarization, and contrast normalization. Removing shadows from phone scans significantly improved extraction accuracy more than simply moving to a larger model. Section-level chunking with rigid schema verification We ceased asking the model to parse entire documents at once. Instead, we chunk the document by specific sections (e.g., Schedule of Payments, Maintenance Obligations) and enforce strict Pydantic schemas. If a field fails validation or deviates from expected parameters, the pipeline flags it for manual review instead of attempting to guess. Treating tables and narrative text as distinct challenges Vision models excel at answering qualitative questions (under what conditions is the deposit refundable?), yet they are notoriously inaccurate with financial tables. We split table extraction into dedicated passes with row-by-row verification. The main takeaway If you are developing with LLMs on messy, physically-generated paperwork, raw context window size will not compensate for the deficiencies. The majority of reliability stems from the cleanliness of the images and the strict schema boundaries. I am keen to share insights if anyone else in the community is grappling with noisy scans and table extraction.",
  "summary": "When we started building the data pipeline for NovAsia (a property catalog and market tool for Southeast Asia), the plan looked embarrassingly simple on paper: Grab property PDFs and developer contracts (SPAs). Run them through a modern vision LLM. Dump structured JSON into PostgreSQL. Call it a day. Two weeks in, the pipeline was completely broken. Here is why standard tutorial advice fell apart…",
  "key_points": [],
  "editors_take": "Relying on raw model context window size fails in emerging markets where low-quality scans and physical stamps necessitate aggressive image preprocessing and rigid schema verification to prevent confident hallucinations.",
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}