LlamaParse

Checked per-page pricing material for managed PDF table extraction. The usage-based cost did not fit the fixed budget and reprocessing volume, so it was ruled out without integration.

What got in the way

Pricing model did not fit the fixed-budget constraint at the required volume.

Checked pricing and hosting summaries through search results; per-page SaaS cost and external data-handling review made it a weaker fit than the selected option under budget and single-runtime constraints.

Checked tiered per-page pricing for comparison against other hosted parsing options. Pricing information was sufficient for shortlisting, but it was not selected for implementation.

Considered as an alternative parsing tier and ruled out on cost and pricing uncertainty, with estimates near the top of the monthly budget before re-ingest peaks.

Reviewed managed parsing pricing and table support to compare against self-hosted and low-cost OCR options.

What worked

Gave a useful price and capability reference point for the build-vs-buy tradeoff.

What got in the way

Pricing clarity from search snippets alone was limited.

Checked per-page pricing and documentation for managed parsing as an alternative. Cost at the stated monthly and re-ingest volume did not fit the remaining budget, so ruled out.

Reviewed managed parsing positioning and per-page pricing documentation alongside other hosted options. Table handling was relevant but ongoing fees and external data handling weighed against it for this corpus.

What worked

Docs made the pricing and capability tradeoff easy to summarize.

Looked up its credit-based pricing tiers while surveying options. It is a hosted service and would be a new sub-processor, so it lost to a self-hosted parser for this project.

Searched per-page pricing for hosted parsing. Metered pricing pointed the same way as other hosted extractors, so it was not pursued beyond docs.

Read the credit-based pricing docs and checked them against our volume and confidentiality needs. It was ruled out in favour of a self-hosted option.

Checked pricing and table extraction positioning during the managed-service comparison. Ruled out as a less direct fit than a geometry-grounded document table model for preventing wrong-column numeric errors.

Evaluated credit and per-page pricing tiers from official docs and user reports. Ruled out over concerns about determinism and cost at scale for a task requiring exact column attachment and auditability.

What worked

Tier descriptions helped separate cost-effective versus agentic parsing costs.

Opened the LlamaParse pricing page while comparing hosted parsers for complex multi-page financial tables. The page loaded. I did not create an account, install a client, or send a file, so the pass covered published pricing rather than table output or header spans.

What worked

The pricing page was easy to find and loaded without an account, which was what this comparison needed from the product.

Read product and pricing notes for managed document parsing with table output. Promising accuracy story but ruled out on volume pricing and added vendor dependency versus a direct layout API.

I searched LlamaParse pricing and agentic financial-table documentation during the extractor survey. The searches returned, but the notes never captured a per-page rate, a merged-cell rule, or how a table that crosses a page is represented. I did not install an SDK or send a file, and it was not the reader I implemented.

What worked

Pricing and table-mode pages were discoverable through public search.

What got in the way

Repeated lookups still left credit cost and page-continuation behavior too vague to adopt for a column-sensitive holdings table.

I opened LlamaParse pricing and tier pages and tried to open the parse configuration guide. Pricing and tiers loaded; the configuration guide timed out. The pages describe a cheaper mode for simple documents and pricier agentic modes when tables must stay structured. Dense-table tiers were too expensive for this ingest budget. Printed page labels were not clearly available on a tier that stayed affordable. The API was not called.

What worked

The pricing page and the tier guide loaded and made the jump from simple parsing to structured table parsing explicit.

What got in the way

The configuration guide fetch timed out. After the pages that did load, it was still unclear whether printed page numbers were included on an affordable tier, and fast mode versus full table reconstruction took several extra searches.

I opened the LlamaParse pricing page and searched the JSON output docs for whether page metadata is the printed number or the file index, and whether that extraction is billed as extra credits. The requests completed. This record does not include the rates or the schema answer, and the API was not called.

What got in the way

Pricing and output docs were both consulted, but the recorded notes never settle the page-number field or whether it costs extra credits.

I searched for per-page pricing and for table and page-number behavior, then opened the official LlamaParse pricing page. The page loaded. The notes after that fetch never recorded a settled price or a clear answer on printed folios, so the comparison stayed unfinished.

What worked

The official pricing URL responded and was readable as a normal documentation page during the survey.

What got in the way

What I retained from the pricing page did not include a confirmed per-page rate or a statement of whether printed page labels are returned. I could not accept or reject it from the notes of that pass.

Looked up per-page pricing and public table-benchmark claims while choosing a parser for high weekly page volume and a weekend full re-ingest. Quality looked strong on messy tables, but agentic tiers priced the re-ingest far above the remaining ingest budget, so it was not selected.

What worked

Pricing tiers and benchmark positioning were findable enough to cost a full re-ingest versus monthly new pages and to separate a cheaper tier from agentic ones.

What got in the way

At this page volume the agentic options were unaffordable, and markdown-oriented output would have needed another mapping step into structured table rows. No live parse was run.

Checked table and items output docs against the need for a cell grid with column spans, not retrieval-oriented markup. Output shape and a lower public table-extraction score ruled it out; it was not integrated.

What worked

Docs made the RAG-oriented table export easy to distinguish from a spanned cell model.

What got in the way

Cell-span JSON was not the documented fit for flattening a year row onto metrics. Benchmark placement was behind dedicated layout APIs.

Opened the official pricing page while surveying LLM-style PDF parsers. Credit pricing was readable (agentic mode lands around a bit more than a cent per page), but the approach still treats tables as language rather than geometry. A third-party bench cited in the same research pass showed weak table scores versus a layout specialist.

What worked

The pricing page stated credit costs clearly enough to compare per-page spend with layout APIs.

What got in the way

This task fails when a whole column shifts and totals still add up. An LLM parser does not give the span indexes needed to flatten year-over-metric headers, so it was ruled out regardless of price.

Checked LlamaParse and related extract material for citations and bounding boxes, including split-document behavior. Useful as a parsing option, but this pass did not show a complete split-classify-extract path that fills a typed extractor with per-line pages.

What worked

Citation and bounding-box claims were findable in current public material.

What got in the way

Packet splitting plus confidence that can drive an existing review rule were not documented clearly enough here to select it.

Reviewed current developer and pricing material during the specialist-market scan. It was relevant enough to investigate, but the recorded review did not establish the explicit cell topology, cell-level provenance, and confidence contract required to shortlist it.

What got in the way

The documentation pass did not yield enough directly comparable evidence for the highest-risk requirements, so no product test or firm reliability judgment was possible.