Datalab’s lift is a 9 billion‑parameter vision model that turns PDFs or images into JSON that follows a user‑provided schema. The core value is that the output is guaranteed to match the requested shape, eliminating the need for post‑hoc validation of structure.
A common pain point is hallucinated fields that look correct but are wrong, especially when data is missing. lift handles this by allowing every scalar field to be null, so absent values are returned as null instead of invented guesses. This gives teams a clear signal to route uncertain results to human review.
Another issue is dealing with multi‑page values such as totals that appear on different sheets. lift reads the whole document in one pass and stitches together cross‑page information, so users do not have to split files or run multiple extraction steps.
Performance matters for production pipelines. lift runs at roughly 9.5 seconds per document on a single GPU, which is several times faster than many hosted alternatives while keeping field accuracy above 90 percent on a curated benchmark. For teams that need per‑field confidence scores or citations, the hosted Datalab API offers those features; for those prioritizing data residency, cost control, or low‑latency self‑service, the open weights can be deployed locally with vLLM.
Schema design is straightforward: use plain JSON Schema, add descriptions for ambiguous fields, keep the schema within the supported subset (no enum, anyOf, $ref, or additionalProperties), and validate the returned JSON against the schema as a safety check. Mark only truly required fields as required so genuine gaps appear as null rather than errors.
A practical workflow looks like: define schema, run lift extraction, check for nulls or errors, send uncertain items to review, validate the JSON, then store or forward the clean data. Re‑using the inference manager across calls amortizes model load, and limiting page ranges helps with very long files.
By combining schema‑constrained decoding with built‑in abstention, lift solves the two biggest extraction problems—invalid structure and false positives—while staying fast enough for batch processing and flexible enough for both self‑hosted and hosted scenarios.
#AI #Product #DataEngineering #Automation #MachineLearning #OCR