Interactive side-by-side viewer comparing LlamaParse v2 parsing tiers on real financial documents.

Live demo: viewer.ts endpoint

Rendering mermaid diagram...

Rendering mermaid diagram...

Document	Pages	Type	Challenge
JPMorgan Q4 2025 Earnings Presentation	21	Slide deck	Charts, graphs, mixed text/image layouts
JPMorgan Q4 2025 Earnings Supplement	29	Financial tables	Dense nested tables, footnotes, merged cells

Tier	Credits/pg	$/page	Best For
fast	1	$0.00125	Quick text extraction, no AI
cost_effective	3	$0.00375	Standard docs, good accuracy/cost ratio
agentic	10	$0.01250	Complex layouts needing reasoning
agentic_plus	45	$0.05625	Maximum accuracy on dense financial docs

(1,000 credits = $1.25 in North America)

JPMorgan Presentation (21 pages × 4 tiers):

Tier	Credits	Cost	Output
fast	21	$0.026	55,097 chars (spatial text only)
cost_effective	63	$0.079	38,572 chars (markdown)
agentic	210	$0.263	38,683 chars (markdown)
agentic_plus	945	$1.181	38,646 chars (markdown)

JPMorgan Supplement (29 pages × 4 tiers):

Tier	Credits	Cost	Output
fast	29	$0.036	213,067 chars (spatial text only)
cost_effective	87	$0.109	186,036 chars (markdown)
agentic	290	$0.363	184,084 chars (markdown)
agentic_plus	1,305	$1.631	178,615 chars (markdown)

Total for this demo: 2,950 credits = $3.69

• fast tier returns spatial text (no markdown structure), hence higher char counts from whitespace
• cost_effective vs agentic produce nearly identical output on these docs
• agentic_plus at 45 credits/page is 15× more expensive than cost_effective with marginal improvement
• For standard financial tables: cost_effective is the sweet spot
• For charts/images: agentic tiers may extract chart data more intelligently

File	Type	Description
dispatch.ts	Script	Sends PDFs to LlamaParse at 4 tiers, saves job manifest
collect.ts	Script	Polls jobs, saves parsed markdown to blob storage
viewer.ts	HTTP	Side-by-side PDF/markdown viewer with cost display
check-pdfs.ts	Script	Utility to verify uploaded PDFs in blob storage

Dispatch — POST /api/v2/parse/

{ "source_url": "https://...", "tier": "cost_effective", "version": "latest" }

Poll/Collect — GET /api/v2/parse/:job_id?expand=markdown

• Response: data.job.status → PENDING | COMPLETED | FAILED
• Content: data.markdown.pages[].markdown (sorted by page_number)
• Fast tier: use ?expand=text → data.text.pages[].text

Key differences from v1:

• Tier-based instead of mode-based
• source_url parameter for public PDFs (no upload needed)
• Response structure: data.job.status not data.status
• 1-based page indexing (v1 was 0-based)