Databricks Improves Document Extraction by 7 Points, Cuts Costs 5–7x

Databricks has added Precision Mode to its `ai_extract` API, targeting three failure modes in LLM-based extraction: cross-page reference resolution in long documents, nested outputs that exceed context limits, and schemas requiring reasoning across multiple sections. The release includes general availability for `ai_parse_document` and a composable SQL-native pipeline — `ai_parse_document` → `ai_extract` → `ai_classify` — that processes documents inside Unity Catalog without external OCR or extraction services. Intercontinental Exchange, Panasonic, and EY-Parthenon already process millions of documents weekly through the system.

Across roughly 9,000 documents spanning 10 internal datasets and 5 public benchmarks (VAREX, RealDocBench, LongExtractBench, LEDGER, Caselaw Access Project), Precision Mode outperforms the next-best frontier model by 7 points. The baseline is chunk-and-merge with GPT, Claude, and Gemini—not single-call processing, which fails on documents exceeding context windows. Test documents ran to 2,000 pages with schemas including 300+ nested fields.

The architecture couples custom, task-specific extraction models with an agentic harness inspired by MemEx. The harness semantically decomposes large extraction jobs, spawns parallel subagents for each chunk, preserves intermediate results, and reconciles them into structured output. This decompose-execute-merge pattern handles cross-page reconciliation: a lease with page-1 renewal terms referencing a clause on page 80 is processed by an agent tracking that dependency rather than page-by-page.

Databricks separates from competitors like Azure Document Intelligence by eliminating redundant reprocessing. VLM-based extraction re-processes the full document on each call. Databricks' parse-once architecture produces a reusable silver layer in Delta; subsequent `ai_extract` and `ai_classify` calls read from that layer. Databricks reports 5–7x lower cost than comparable pipelines and 3–5x lower cost versus leading VLMs on OmniOCR benchmarks. Loopback Analytics, processing clinical notes, achieved 90% lower cost than its prior approach.

The quality gain extends to RAG. Pre-processing documents through `ai_parse_document` before ingestion into agent frameworks produced a 16% average performance gain on treasury bond documents via OfficeQA. Frontier agents without better parsing scored below 50% on document-reasoning tasks—not from weak reasoning models, but from malformed input text. Tables rendered as flat text, cross-column references broken, figure data lost. Fixing the extraction layer, not the reasoning layer, moved the number.

For RAG teams: the pipeline is SQL-native. `ai_parse_document` ingests PDFs, images, Word docs, and slides via Lakeflow Spark Declarative Pipelines, writing structured JSON to Unity Catalog. Downstream `ai_extract` calls reference that output. The same SQL call handles 100 invoices or 100,000 on serverless batch infrastructure without rearchitecting. Unity Catalog provides lineage, fine-grained access control, and auditability.

One limitation: `ai_parse_document` is generative, not deterministic. Community testing shows occasional transcription inaccuracies on precise identifiers—names, IDs, codes—where deterministic parsers would be exact. For workflows prioritizing identifier precision over layout fidelity, evaluate a hybrid approach (deterministic parse for machine-readable PDFs, `ai_parse_document` for scanned or layout-heavy documents) before committing.

If your RAG pipeline has a mediocre extraction step, fixing it upstream with `ai_parse_document` and Precision Mode is now the highest-leverage accuracy improvement available in the Databricks stack at production scale.