AI · Document AI
How AI reads a lighting specification: from PDF to a structured fixture schedule
A fixture schedule looks like a table. Inside a PDF it is a scatter of text boxes, merged cells and footnotes. Turning it into data is the first job AI has to get right.
Short answer
AI reads a lighting specification in five steps: layout analysis (find pages, sections and tables, with OCR for scans), table extraction (rebuild the fixture schedule’s rows and columns), entity extraction (type marks, manufacturers, catalogue numbers, lumens, CCT, CRI, voltage, dimming, mounting) using language models, normalization (units, lumen per foot vs per fixture, option codes), and validation (cross-checks against the plans and the spec text, with low-confidence lines sent to a person). The output is a structured schedule that can be matched, priced and audited.
Key takeaways
- Fixture schedules are tables in appearance only; PDFs store them as loose text.
- Language models are good at extracting attributes but must be checked.
- Normalization (units, per-foot vs per-fixture) prevents silent errors.
- Cross-checking against plans and spec text catches what extraction misses.
- Low-confidence lines go to a person instead of guessing.
Why fixture schedules are hard for machines
To a person, a fixture schedule is an obvious table. Inside a PDF there is no table: there are text fragments with coordinates, lines drawn separately, merged header cells, wrapped descriptions and footnotes. Scanned sets add OCR noise. Schedules also refer elsewhere: quantities live on the plans, requirements in the spec section, and changes in addenda.
The five-step pipeline
1. Layout analysis
Find the document’s structure: sections (for example the interior lighting section of the specification), pages with schedules, title blocks and notes. Scanned pages go through OCR or an OCR-free document model.
2. Table extraction
Rebuild rows and columns from the fragments: detect the header, resolve merged cells, join wrapped lines, and keep footnote markers attached to their rows. This step is where most copy-paste errors come from, and where specialised table models help.
3. Entity extraction
Language models read each row and its description and pull out attributes into a fixed schema:
| Field | Example |
|---|---|
| Type mark | L1 |
| Manufacturer / catalogue no. | Brand X LIN2-R-40K-XX-0-10V |
| Output | 400 lm/ft |
| CCT / CRI | 4000K / 90 |
| Voltage | 120–277V |
| Dimming | 0-10V |
| Mounting | Recessed, drywall, trimless |
| Emergency | Battery on 1 in 4 |
The schema matters: a model asked to “summarize the table” invents structure; a model asked to fill defined fields with evidence can be checked.
4. Normalization
Turn what was written into comparable values: lumens per foot vs per fixture, “4K” vs “4000K”, “0-10V dimming” vs “1-10V”, open catalogue options (“XX”) flagged as decisions, not values.
5. Validation and review
Cross-check the extracted schedule against the plans (type marks present, quantities, linear lengths), the spec text (substitution rules, required certifications) and internal logic (a 347V project with 120–277V drivers is a flag). Each line carries a confidence; low-confidence lines go to a person. This is the difference between a demo and a system you can quote from.
What research says
Academic work on construction specifications shows large language models can structure these documents well but not perfectly. One 2024 study converting specification tables of contents to JSON reported 85% accuracy with an OCR-free document model and 89% with GPT-3.5 Turbo. Benchmarks on building structured tables from documents find even strong models struggle on complex layouts. The practical lesson is consistent: extraction plus validation plus review beats extraction alone.
Why it matters for procurement
Every later step depends on this one. A wrong voltage or a missed open option becomes a wrong quote, a rejected submittal or a wrong product on site. Good extraction makes product matching possible and makes errors visible early. For how people read the same schedule, see how to read a lighting fixture schedule.
Send a schedule and SuppLux’s AI matches every line to verified products, with the best alternative shown first, then turns the matched set into a quote and submittal package. Send a schedule.
Frequently asked questions
Can AI read scanned lighting specifications?
Yes, with OCR or OCR-free document models. Scans add errors (misread characters in catalogue numbers, broken table lines), so confidence scoring and review matter more for scanned documents.
How accurate is AI extraction from construction specifications?
It depends on document quality and the task. Published research on construction specifications reports accuracy in the high 80s percent for structuring sections with large language models; production systems add validation and review to close the gap.
What does AI extract from a fixture schedule?
Type marks, descriptions, manufacturers and catalogue numbers, lumen output, wattage, CCT, CRI, voltage, dimming protocol, mounting, finish, emergency options and remarks, plus the links to plan quantities and spec requirements.
Why not just copy the table into a spreadsheet?
Copy-paste from PDF breaks merged cells and wrapped text, and it cannot resolve open options or units. Extraction plus validation produces data that can be matched and audited.
References
Keep reading
- AI · Roadmap AI in lighting procurement: a technology roadmap from spec to site A technology roadmap for AI in commercial lighting procurement: reading specifications, matching every line to verified products, assembling quotes and submittals, and the next stages, learning from review, supply-chain intelligence and agentic workflows. Where AI helps today, where people stay in the loop, and what to measure. →
- Specification · Drawings How to read a lighting fixture schedule (and what is usually missing) A field guide to the lighting fixture (luminaire) schedule on electrical drawings: what each column means, how it ties to plans and the specification, and the details most often missing that cause wrong quotes and rejected submittals. →
- AI · Product matching AI product matching for lighting: how an equal is found and ranked How AI product matching works for commercial lighting: why it is an entity-resolution problem, how candidates are generated and filtered by hard constraints (voltage, dimming, mounting, certification), how equals are scored on photometrics and colour, and why every match needs evidence and a human decision. →
- Platform · Spec matching AI spec matching for commercial lighting SuppLux matches every line of a lighting specification to verified products and ranks equal or better alternates by photometrics, electrical data, mounting and lead time, with a specialist review before quoting. →