中文
AI Engineer World's Fair

An open source library that turns PDFs into structure an LLM can actually understand — and lets you drop the vector database

Structuring the Unstructured - Cedric Clyburn, Red Hat · Cedric Clyburn

21 min
ContextAgent

21 min total·Actually worth watching closely: ~8 min·3 must-watch clips

Orange = the 8 minutes worth watchingFor the rest, the guide is enough
Segment guide · 8 segments
  1. 0:00 2:40Listen

    Context is the bottleneck

    The opening claim: what caps an AI application or agent is the context you feed the model, and most of the valuable data inside an enterprise is locked in PDFs, presentations, contracts, scanned documents and diagrams the model simply can't read.

    Unstructured data is becoming the new context layer for AI, and today's options are either proprietary services or require handing private data to a third party

    Mostly the speaker making the argument standing up, with only a few lines of title text on screen — fine to listen to while doing something else▶ Jump to 0:00
    Speaker · Cedric Clyburn
  2. 2:40 4:35Listen

    A term that was never real

    A true incident: reading an old scanned two-column PDF, an AI merged two words from two different columns into a term that doesn't exist — and that word now appears in 20 scientific papers and is being cited by others.

    A parsing error isn't a formatting problem, it's knowledge pollution that spreads down the citation chain

    Pure storytelling, no visuals — the single most retellable stretch of the talk▶ Jump to 2:40
    Speaker · Cedric Clyburn
  3. 4:35 7:03Skim

    Neither path is good

    A comparison of the two extremes: simple parsers are fast and cheap but the output is dire; frontier models parse well but run around $30 per million output tokens, give different results each time, and change format whenever the model version moves.

    Cost isn't the worst problem with the frontier-model route — non-deterministic output and hallucinations at scale are

    The screen shows side-by-side parsing output and cost figures; one glance at that garbled export is enough, no need to watch frame by frame▶ Jump to 4:35
    Speaker · Cedric Clyburn
  4. 7:03 9:37Skim

    Docling is the middle ground

    An introduction to this Linux Foundation open source project: it combines OCR with layout analysis to convert documents on local CPU into Markdown or JSON that keeps the original page layout structure, and extends to captioning images automatically with vision models.

    It runs on local CPU, which fits teams that don't want to pay for a service or sit inside a fully air-gapped network

    Backed by a capability overview and a document structure diagram — skim the module names on the slide and focus on the middle stretch about what image annotation is for▶ Jump to 7:03
    Speaker · Cedric Clyburn
  5. 9:37 12:09Watch

    Three lines to convert a paper

    Cut to the live demo: after the install, a converter object takes in an 8-page PDF and returns a typed data object you can query for page count and table distribution, then export as Markdown or a web page.

    The conversion result isn't a blob of text — it's a typed document object you can query programmatically

    This is where the hands-on portion starts, with code and output scrolling on screen; follow along and you can copy the usage directly▶ Jump to 9:37
    Speaker · Cedric Clyburn
  6. 12:09 14:03Watch

    Tables, images and bounding boxes

    Finer-grained extraction: eight tables exported as queryable data; the pipeline adjusted to scale up and extract images, yielding a mapping of source image, caption and embedded text elements; and finally a visualizer drawing boxes around every extractable element on the page.

    You can also extract just the fields you name — an invoice's bill number and total price — without converting the whole document

    The bounding box segment lays out exactly which blocks on the page the model understood, the densest visual in the talk — the narration alone doesn't convey it▶ Jump to 12:09
    Speaker · Cedric Clyburn
  7. 14:03 16:40Watch

    Throw out the vector database

    A counterintuitive retrieval approach: no chunking, no embeddings, no vector database — just use the generated document outline (a summary per section) as the index and let the model pick the relevant section in an agentic loop, pull the full text and answer. Demoed end to end.

    The approach still holds on an annual report with 418 sections, handling complex questions through multiple iterations

    You can watch the model pick sections and pull source text round by round, which is where this segment's persuasiveness lives; hearing only the conclusion sounds like a boast▶ Jump to 14:03
    Speaker · Cedric Clyburn
  8. 16:40 20:40Listen

    Scaling up and wiring into agents

    The closing engineering section: Hugging Face used it to clean web-crawled PDFs for the FinePDFs dataset at roughly 50x the cost savings over vision models and OCR directly, and on CPU; for hundreds of thousands of files you can deploy it as a REST microservice, or expose the tools to coding assistants over a standard protocol.

    Once it's wired into an agent you no longer memorize arguments and commands — one sentence gets conversion, summaries and cross-file roundups done

    Mostly architecture explanation backed by numbers; the closing demo of a local model with the editor is short and easy to follow by ear▶ Jump to 16:40
    Speaker · Cedric Clyburn