Watch the video with chapters & transcript →
DATA ENGINEERING
In short: A Claude Cowork project answered questions correctly across five file formats — PDF, DOCX, XLSX, CSV, and JPEG — with no embeddings, no vector database, and no code.
A traditional retrieval-augmented generation pipeline means chunking documents, generating embeddings, running a vector database, and maintaining all of it as files change. Ani Björkström, a tech consultant based in Stockholm, tested whether Claude Cowork projects can deliver the same result with none of that infrastructure, using a folder of mixed-format finance files.
Projects accept much larger document sets, because a normal chat has a smaller limit on document sizes. A project also connects to a folder path on your computer, either created from scratch or pointed at an existing catalog, so every file in that folder becomes queryable without individual uploads.
The test project was connected to an existing folder containing five files: a PDF annual report, a DOCX professional services agreement, an XLSX quarterly sales report, a CSV transaction log, and a JPEG invoice. Creating the project and linking the folder takes a few clicks, with no ingestion step to configure.
Claude answered all five test questions correctly, one per file format, without ever being told which document held the answer. In each case it identified the right file on its own and cited where the figure came from, down to the page, section, or sheet.
| File format | Question asked | Claude's answer (verified correct) |
|---|---|---|
| PDF annual report | Net income for 2025 | $612 million, cited from page 5 |
| DOCX services agreement | Termination terms | 60 days prior written notice, per section 2.2 |
| XLSX quarterly sales data | Monthly active users for 2025 | 892,000, from the key KPI sheet |
| CSV transaction log | Sum of the training category | $106,005.65 across 48 rows |
| JPEG invoice | Invoice due date | November 14, 2025 |
The CSV case stands out: Claude aggregated an amount column filtered by category and reported the 48-row count, which matched a manual check of the file. The JPEG invoice shows images work too, with the November 14, 2025 due date read correctly.
For a document set like this, no — Claude retrieved correct information from every format without any RAG infrastructure to build or maintain. That removes the ongoing maintenance burden of a traditional pipeline: no re-chunking when documents change, no embedding model to manage, no vector store to operate.
For finance teams, this means a controller or analyst can point a project at a folder of reports, contracts, spreadsheets, and scanned invoices and start asking questions in minutes.
No. In all five tests Claude scanned the connected folder, picked the correct document, and cited the location — for example page 5 of the annual report or section 2.2 of the agreement.
Yes. It read a due date of November 14, 2025 from a JPEG invoice, pulled 892,000 monthly active users from an XLSX KPI sheet, and summed $106,005.65 across 48 CSV rows.
Chat has a smaller limit on document sizes, while a project connects to a whole folder on your computer. That makes projects the practical choice for document sets of any real size.
Read the full video transcript, with chapters, on the video page →
Want this working inside your finance team?