Llamaparse And Llamacloud

An integrator uses LlamaParse fast mode on complex financial PDFs with nested tables, then complains tables are flattened and misaligned in the parsed markdown.

Select the LlamaParse parse mode for the document complexity — higher-fidelity modes (e.g. accurate / premium / multimodal) for complex tables and layouts, fast mode for simple text — trading cost/latency for fidelity. Validate parsed output on representative documents before bulk processing.

LlamaParse parsing is asynchronous (submit then poll). The integrator reads the result immediately after submit and treats a not-yet-ready job as a parse failure.

Treat parsing as an async job: submit, then poll job status with backoff until it reaches a terminal SUCCESS/ERROR state (or use the SDK's blocking helper) before reading results. Distinguish 'still processing' from 'failed' and handle ERROR jobs explicitly.

The LlamaCloud API key is embedded in client-side code / a notebook shared publicly so LlamaParse calls work from the browser.

Keep LLAMA_CLOUD_API_KEY server-side, loaded from env/secret manager, never in client code or shared notebooks. Proxy LlamaParse/LlamaCloud calls through a backend. Rotate the key if exposed; a leaked key allows unauthorized billable parsing.