Report Insight Lens
Status: Live.
Your muse
Rebuild ofLlama Best Practices: Unlock Rich Insights from Rich Documents (opens in a new tab)Clean
Upload a PDF report and pull out the key numbers, tables and charts. This rebuilds Meta's video "Llama Best Practices: Unlock Rich Insights from Rich Documents" (https://dev.meta.ai/resources/videos/unlock-rich-insights-from-rich-documents) on current Muse models. It demonstrates: Extracting data from PDFs, invoices and reports with the DocumentLens recipe. Plan: Upload a PDF report. Spark 1.3 reads it directly and answers with tables and page references, and SAM 3.1 outlines the charts and figures on page images.
Checks passed
The plan: Report Insight Lens
Upload a PDF report and pull key numbers, tables and charts.
- Upload a PDF report or invoice
- View auto summary with key numbers and extracted tables
- Ask a question or pick a chart label to locate
- See answer with tables and page references plus outlined figures
See the full plan
Who it's for
Analysts, founders and students working with reports and invoices.
How it works
- Upload a PDF report or invoice
- View auto summary with key numbers and extracted tables
- Ask a question or pick a chart label to locate
- See answer with tables and page references plus outlined figures
Screens
- Upload report. Drop a PDF and start extraction with progress.
- Insights dashboard. Show summary, key numbers and extracted tables with page refs.
- Ask and locate. Ask questions and view outlined charts and figures on pages.
What goes in
- PDF: Report, invoice or paper up to request size limit. Text based PDF preferred.
- Text (optional): Optional question about numbers, tables or trends. Under 500 characters.
What comes out
- Text: Summary with key numbers plus tables and page references for each fact.
- Overlay: Boxes outlining charts, tables and figures over rendered page images.
Models
Model Used for Why Muse Spark 1.3 Read the PDF directly and extract numbers, tables and Q and A answers with page references. Spark reads PDFs and returns structured text and JSON with reasoning and vision. SAM 3.1 Outline charts, tables and figures on rendered page images. SAM finds and boxes objects from a short noun prompt on images. Limits
- Rebuild fit: clean. Same demo as the video rebuilt on Spark 1.3 and SAM 3.1 without original footage, script or Llama code.
- Large or scanned PDFs over about 4.5 MB must be shortened or compressed before upload.
- Answers use only the uploaded PDF. No web search or outside data.
- Figure boxes need simple prompts like chart or table. SAM can miss small or overlapping figures.
- Saves stay in browser localStorage. No accounts or database.
From your inspo
Rebuild of Llama Best Practices: Unlock Rich Insights from Rich Documents. Keeps the upload a PDF report flow to pull key numbers, tables and charts with page references on current models.
Takes
- Extracting key numbers and tables from PDF reports and invoices
- Answering document questions with tables and page references
- Outlining charts and figures on page images
Leaves out
- Video footage and script
- Meta and Llama names and logos
- DocumentLens recipe code and exact look
Cost per run: about $0.08 (est.). Spark read of 50k input tokens at $1.25 per M equals $0.0625 plus 2000 output tokens at $4.25 per M equals $0.0085. SAM on 3 page images at $0.002 each equals $0.006. Total $0.08.
Recorded demo Done.
npm run verifyexits 0 — all 6 steps pass, 4 tests green, manifest valid.