Use case · Multimodal RAG
Document and PDF Q&A
Real documents are more than text. Answers often sit in a table, a chart, or a scanned page, and the agent has to find them and cite where they came from.
The experiment
Parse real-world PDFs into text, tables, and images, index them, and compare retrieval and generation setups side by side.
- Dataset
- MMDocIR, a benchmark of long, real-world PDF documents with text, tables, and figures.
- Agent
- A multimodal RAG pipeline that parses each PDF, indexes its contents in Pinecone, and answers from the retrieved pages.
- Knobs tested
- Parsing strategyChunkingEmbedding modelRetrieval depthGenerator model
What the experiment showed
- 01Runs end to end in one notebook, so you can reproduce it before changing anything.
- 02Every stage, from parsing to generation, can be varied in the same experiment.
How RapidFire helped
Every configuration in one run
Compared the knobs side by side on the same hardware, instead of one run after another.
Live results
Scored each version on slices of the eval set as it ran, so decisions came early.
Weak runs stopped early
Cut directions that were not working and put the compute behind the ones that were.
Winners cloned
Cloned the strongest run and changed one factor at a time to find what drove the gain.
Apply it to your data
- 1Point the parser at your own PDFs: reports, manuals, contracts, or filings.
- 2Test parsing and retrieval choices side by side.
- 3Pick the setup that finds answers in tables and figures as well as text.