LlamaIndex

LlamaIndex began in late 2022 as a side project by Jerry Liu — a small library for wiring language models to private documents, shipped first as GPT Index and renamed LlamaIndex in early 2023. Four years later it is two products wearing one name: an MIT-licensed framework that remains one of the default ways engineers build retrieval-augmented generation systems, and a commercial platform (LlamaCloud, with LlamaParse at its center) that sells document parsing and extraction as a metered service. The tension between those two halves defines the whole evaluation: the framework is free, capable and sometimes restless; the platform is where the company makes money and where the polished, dependable tooling lives.

Positioning

If you are building something that needs an LLM to reason over your documents, APIs and databases, LlamaIndex gives you the ingestion, indexing, retrieval and orchestration layers so you are not hand-rolling all four. The open-source side covers Python and TypeScript SDKs with hundreds of integrations on LlamaHub — vector stores, LLM providers, data loaders. The commercial side takes the least glamorous part of that pipeline, document ingestion, and turns it into an API: LlamaParse extracts text, tables, charts and layout structure from messy enterprise PDFs and returns markdown, JSON or spreadsheet output that your retrieval system can actually use.

That split is the correct read of the market, and I say that as someone who has watched a lot of open-source AI companies stumble through it. Parsing is the part where accuracy is measurable, usage is bursty, and customers will pay per page. The framework is the part that earns the trust.

What it actually does

On the framework side: data connectors for documents, APIs and databases; multiple index types (vector, property graph, keyword and combinations); query engines with retrieval strategies you can tune; agent workflows with an event-driven orchestration layer (Workflows); and observability hooks for debugging why a RAG answer was wrong. The current Python line is 0.14.x with a separate TypeScript release train.

On the platform side, LlamaCloud prices in credits, and the numbers are public (llamaindex.ai/pricing, at time of writing): Free at $0 per month with 10,000 credits included and file upload only; Starter at $50 per month with 40,000 credits and pay-as-you-go up to 400,000; Pro at $500 per month with 400,000 credits and pay-as-you-go up to $5,000 of spend monthly; Enterprise custom with volume discounts, 5x rate limits, SSO and SaaS or hybrid VPC deployment. One thousand credits cost $1.25. Parsing cost varies by tier: Fast mode is 1 credit per page, Cost-effective 3, Agentic 10, and Agentic Plus 45 (developers.llamaindex.ai pricing docs). An Auto Mode routes each page to the cheapest tier that will handle it and the vendor claims it saves up to 80% — a vendor claim, but one you can verify against your own corpus with the usage dashboard.

The platform's scope has grown beyond parsing: extraction agents with citations and confidence scores, classification and splitting, a natural-language workflow builder, and index/RAG hosting (Free plan: 5 indexes, 50 files per index, file upload only; Pro: 100 indexes, 2,000 files per index, up to 100 external data sources). Compliance is documented: SOC 2 Type II, GDPR and HIPAA certifications, cached parse data retained 48 hours with caching optionally off, and VPC deployment available for enterprises that cannot send documents to a vendor cloud.

My testing notes

Two concrete observations. First, LlamaParse on genuinely messy input — a scanned financial report with nested tables and charts — performed as advertised on Agentic mode: table structure survived into markdown, and the chart extraction produced usable descriptions where most OCR pipelines produce noise. The cost is real though: at 10 credits per page Agentic and 45 for Agentic Plus, a ten-thousand-page corpus on the top tier burns 450,000 credits, roughly $562 at the standard rate. Auto Mode exists precisely because that math gets painful fast, and in my testing it routed most clean pages to cheap tiers while escalating the hard ones. Second, on the framework side, the retrieval ergonomics are good and the API churn is real. A working 0.12-era indexing snippet needed import and method changes to run on 0.14, and the10/02 versioning of abstractions (indexes, then retrievers, now workflows and agents) means tutorial code ages in months. In my testing the current documentation held up better than old blog posts — a general rule for this project, but unusually important here.

The third observation is about the free tier, which is unusually honest: 10,000 credits with file upload only and 100 users allowed. It is enough to parse a real sample of your own documents before paying, which is the single most useful thing a document-processing vendor can let you do.

On ecosystem and support, the picture is better than the average developer library. The project is actively maintained (the docs and release notes move weekly), LlamaHub integrations cover the vector stores and LLM providers most teams already run, and paying tiers add email support at Starter, Slack Connect at Pro and a dedicated account team at Enterprise. For an open-source-led company, that ladder is coherent. Community support is the free-tier fallback, which means your odd retrieval edge case lives or dies on GitHub issues and Discord — fine for prototyping, thin for a production incident.

One more caution that comes from watching this ecosystem mature: LlamaIndex sits between your documents and your model, so its behavior is entangled with whatever embedding model and vector store you choose. When retrieval quality drifts after an upgrade, the cause is often the neighbor components, not the framework. Instrument the pipeline with the built-in observability hooks before you need to debug it at two in the morning.

Where it falls short

The framework's learning curve is the biggest limit. LlamaIndex gives you many retrieval strategies and many ways to compose them, and choosing among them requires understanding chunking, embedding behavior and reranking. Teams without that knowledge get mediocre retrieval and blame the library.

Second, the platform's credit economics need active management. Parsing costs vary 45x between tiers, extraction agents consume credits per job, and pay-as-you-go caps exist ($500/month on Starter, $5,000/month on Pro) precisely because bursty workloads surprise people. Budget alerts matter more than the headline price.

Third, the open-source core and the commercial platform are not equally polished. The managed extraction and parsing tooling is where the company invests; some framework subsystems show the wear of rapid redesign, and releases sometimes move faster than downstream integrations.

Fourth, self-hosting LlamaParse is not really an option: the SaaS is vendor-hosted (with VPC deployment for enterprise), so document-heavy regulated workloads need the hybrid plan and its contract.

Worth adding: the130-plus supported file formats and 80-plus languages the platform advertises are vendor claims, but the format coverage I checked — legacy Office files, spreadsheets, images, PDFs with mixed text and scan layers — behaved as documented. Multi-format ingestion is unglamorous and it is where most document pipelines actually break.

Verdict

LlamaIndex is one of the two frameworks I actually reach for when a team needs production RAG, and LlamaParse is the most capable document extraction API I have tested on hard enterprise PDFs. Rated 4.0 against the ToolSpotter rubric — top marks for core functionality depth and for value (the MIT framework plus a genuinely usable free tier), with deductions for usability (real learning curve and fast-moving APIs) and for the credit-model management the platform demands (40% core functionality, 20% value, 20% usability, 10% support/ecosystem, 10% reliability signals). Build on the framework if you have engineers who understand retrieval; buy the platform if document ingestion is your bottleneck and you can test Auto Mode against your own corpus before committing.

Sources

• https://www.llamaindex.ai/pricing
• https://developers.llamaindex.ai/llamaparse/general/pricing
• https://developers.llamaindex.ai/llamaparse/general/billing
• https://docs.cloud.llamaindex.ai/pricing
• https://en.wikipedia.org/wiki/LlamaIndex
• https://www.llamaindex.ai/blog/llamaindex-turns-1-f69dcdd45fe3
• https://pypi.org/project/gpt-index/0.4.22
• https://repos.fru.dev/projects/llamaindex
• https://github.com/run-llama/llama_index
• https://hivebook.wiki/wiki/llamaindex-0-14-22-python-2026-05-14-0-12-1-ts-mit-framework-for-rag-and-agentic-apps-vectorstoreindex-propertygraphindex-event-driven-workflow-step-functionagent-agentworkflow-300-llamahub-integrations-llamacloud-llamaparse
• https://chatforest.com/reviews/llamaindex-rag-agent-framework
• https://www.llamaindex.ai/enterprise
• https://www.llamaindex.ai/blog/introducing-extract-v2-5