Retrieval-augmented generation has a dirty secret: most RAG failures aren't retrieval failures, they're parsing failures. If your parser interleaves two columns of text or flattens a nested table into a run of numbers, no embedding model or reranker downstream can recover the meaning. The parser is the load-bearing layer, and it's the one most teams pick last.

Here's the fastest way to choose. The four leading parsers split into two camps, and the split is the whole decision.

The two camps#

VLM-agentic parsers — Reducto and LlamaParse's higher tiers — read a page the way a person does. Layout-aware vision models find the regions; a vision-language model interprets each one in context; and in Reducto's case a multi-pass "agentic OCR" loop re-processes low-confidence blocks until the output is internally consistent. This is what handles a scanned invoice with a rotated, nested table without silently scrambling it.

Layout engines — Docling and Unstructured's open-source library — detect structure with specialized models and rules, no general VLM review pass. Docling runs IBM's DocLayNet for layout and TableFormer for tables, locally, on commodity hardware. Unstructured labels each element (title, paragraph, list, table) with type and position so you can drive chunking off the semantics. They're fast, free, offline — and they don't self-correct.

The camp you need is decided by your worst documents, not your average ones. Parsers all look fine on a clean digital PDF; they diverge on the messy 10%.

The benchmark everyone will quote at you#

You'll see one number thrown around: on LongExtractBench, Reducto Deep Extract scored 99.6% precision and 99.6% recall across 225 long documents, while LlamaIndex's LlamaExtract landed at 80.0% and 77.5%. That gap is real in the published results — but read the provenance. The benchmark was built and run by micro1; the headline reached you through Reducto's own press release, and Reducto won it. Frontier LLMs tested alongside (GPT-5.5, Opus 4.8, Gemini 3.1 Pro) completed only about half the corpus.

Treat it as a signal that the agentic camp handles long, dense documents well — not as a neutral ranking. The only benchmark that decides your pipeline is your own worst 50 documents.

How to actually choose#

Are your documents messy and high-stakes? Financial statements, multi-column legal PDFs, scanned forms where a silent error costs real money — go agentic. Reducto leads on the hard inputs and gives you the deployment options below; LlamaParse's Agentic and Agentic Plus tiers get you there if you're already in the LlamaIndex ecosystem.

Do you need on-prem, air-gapped, or HIPAA? Docling and Unstructured OSS run fully local with zero cloud egress — the pragmatic default for a no-egress shop with clean-enough docs. For managed compliance, Unstructured Platform is SOC 2 Type II and HIPAA with in-VPC deployment, and Reducto offers in-VPC, on-prem, and air-gapped with zero data retention. One trap: Unstructured's free OSS library is not itself certified — that's the commercial Platform.

Cost-sensitive? LlamaParse lets you tune spend per page — Fast at 1 credit, Agentic at 10, Agentic Plus at 45 — so you send only the hard pages to the expensive tier. Docling and Unstructured OSS are free. Reducto's standard plan includes 15,000 free credits and a 20% discount for batch jobs you can wait 12 hours on.

The real risk isn't picking the "wrong" parser — it's picking one, ingesting a million pages, and discovering the errors months later when retrieval quality quietly craters. Re-parsing a corpus is the expensive mistake. Spend an afternoon running your two finalists over your ugliest documents first; the accuracy gap only shows up where it matters.