For RAG, compare retrieved evidence and answer correctness using the same queries. A parser’s framework integration does not establish retrieval quality.

Keep the evaluation controlled

Freeze a representative PDF set and questions with known answers. Convert the same pages with each parser, then use the same chunking, embedding and retrieval settings. Inspect whether the correct passage appears in the top results, whether tables preserve labels and units, and whether answers cite the right source.

LlamaParse documents Markdown, text and JSON outputs through its parsing service. It is not restricted to LlamaIndex-only retrieval. Official parsing documentation, checked 10 September 2026. pdfToMarkdown returns a Markdown string with conversion metadata. Neither choice requires a particular vector database.

Preserve evidence during chunking

Store a stable document identifier and source reference with each chunk. Carry heading context into subsections. Keep table headers with their data and avoid splitting HTML spans as if they were plain pipe rows. Bound long sections using the embedding model’s actual tokenizer.

Do not invent page citations from Markdown line numbers. The pdfToMarkdown response reports the processed page count; it does not promise a page-to-character map.

Retry before re-indexing

Save one conversion identity per immutable input and options. Recover uncertain account work using that identity before issuing a new conversion. Deduplicate the downstream index separately: a replayed API result should not insert a second copy of every chunk.

Measure cost and quality separately

Record delivered pages, total elapsed time, parsing defects and retrieval outcomes. There is no current paired LlamaParse benchmark here. The general comparison covers interfaces; the RAG pipeline guide covers ingestion.

Convert a document

Use the tested Python, JavaScript or curl workflow. Choose one input: a public PDF URL, Base64 bytes, or a raw PDF upload. Raw uploads avoid Base64 expansion; all three send the document to an external processor.

The public demo converts page 1 with a watermark, at 3 requests per minute per IP. New accounts receive 20 trial pages once. Further account conversions use paid credits: monthly pages expire, top-ups do not. The API reference owns authentication, limits, replay and retention.