To automate tender document analysis with AI, push the PDF through a field extractor that returns typed JSON matching the exact schema your system expects - parties, deadlines, penalties, fine amounts, line items. The newest open option is Dharma Field Extractor, a 2-billion-parameter model that takes a document image plus a schema and gives back structured fields in a single call.
Dharma-AI released it on Hugging Face on November 4, built on DharmaOCR-LITE v1.1 with a LoRA adapter trained on 43,384 examples across English, Portuguese and a smaller share of European and Asian languages. Two modes: open (no schema, returns whatever key-value pairs it detects from layout) and schema-directed (your exact keys and types, with null where the document lacks a field). The same checkpoint handles invoices, IDs, medical requisitions and contracts, with no per-type fine-tuning.
For a small business, this is the piece that used to be expensive. Prompting a frontier LLM for every incoming invoice, ID card or tender annex adds up fast at high volume, and prompt drift silently breaks downstream systems. A 2B model you can host cheaply drops into any n8n webhook or Python pipeline - the same way I wire schema-typed output into a tender parsing workflow that pulls penalties and deadlines out of hundreds of pages. Watch recall on messy scans and very long nested tables before you cut the API over.
Why it matters: If you move documents into a database at volume, a small specialized extractor now beats frontier APIs on cost and latency for the exact JSON shape your system already uses.
Source: Hugging Face