> ## Documentation Index > Fetch the complete documentation index at: https://docs.realtimelca.com/llms.txt > Use this file to discover all available pages before exploring further. # Extract From PDF > Let the AI read an EPD document and fill the Create Material form for you. Typing an EPD into the Library by hand means transcribing several hundred numbers across four indicator tables. **Extract From PDF** reads the document instead and fills the form for you, leaving you to check the result rather than key it in. It sits at the top right of the [Create Material](/user-guide/library/materials/create-materials) form, on both the **Published** and **Custom** tabs. ## Run an extraction Click **Extract From PDF** and choose the EPD document. Extracting from PDF dialog explaining that it takes about 15 to 60 seconds, with a cancel button.

Extracting from PDF dialog explaining that it takes about 15 to 60 seconds, with a cancel button.

Extraction takes roughly **15–60 seconds** depending on the document. **Cancel** stops it and leaves the form as it was. ## Check what came back The form comes back populated — name, life time, owner, data source, description, publication and validity dates, EPD type and specification, declared and mass values with their units, and the indicator tables. Create Material form filled in after extraction, with general details, description, EPD specification, and EPD units populated from the document.

Create Material form filled in after extraction, with general details, description, EPD specification, and EPD units populated from the document.

Anything the extraction could not find stays empty and flagged, so the gaps are easy to spot before you save. Work down the form and compare it against the document, paying most attention to: | Check | Why | | ---------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------- | | **Declared value and unit**, **mass value and unit** | Everything downstream is per declared unit. A unit read as *kg* instead of *kg/m³* changes every result the material touches. | | **Valid to** | Decides whether the material is still usable in a compliant calculation. | | **Type** and **Specification** | *Generic* versus *Product*, and +A1 versus +A2, affect how the material is treated and compared. | | **Indicator values** | Spot-check the modules that carry the most weight — A1–A3 above all. | Then **Create Material** as usual. ## How reliable is it? Real-Time LCA published its own testing in [Introducing AI-enabled EPD import](https://community.realtimelca.com/t/introducing-ai-enabled-epd-import/490): across roughly **10,000 values** from **35 test files**, the extraction returned **98% of the values correctly**, with **11 individual entries wrong**. The remaining values were either missing or flagged as potential errors for a human to review. That accuracy comes from a **neuro-symbolic** approach — generative AI does the reading, and symbolic rules reason over the output to catch hallucinations. The rules detect mismatches; they do not remove them. Extraction is an accelerator, not a substitute for checking. A wrong indicator value is invisible once the material is in the Library — it simply produces a wrong answer in every project that maps to it. Review the form against the source document before saving. For how Real-Time LCA handles AI features generally, see [Transparency and privacy](/user-guide/integrations/ai-tools/transparency-and-privacy). ## Next steps What each field on the Published form means, field by field. Find the material once it is saved, and check it in the detail panel.