LlamaIndex, the shortest path from a client's documents to a query engine

Ingestion breadth is why it exists and it is why you reach for it. Connectors for APIs, PDFs, SQL databases and a long tail of formats mean the first two days of an engagement, which are always about getting the client's actual documents into something queryable, become hours instead. Indexes, query engines and chat engines form a sensible ladder: a working retrieval-augmented answer in roughly twenty lines, and then you take the ladder apart piece by piece as the real requirements arrive. Workflows are event-driven rather than chain-shaped, which fits systems that branch, and most useful systems branch. LlamaParse handles the document that always turns up on day three, the scanned PDF with merged cells in a table that matters, and it does so with a vision model rather than a regular expression and hope. The free monthly credit allowance is generous enough to find out whether it helps before anybody has signed anything.

The API has moved enough that the internet is now full of confidently wrong answers. The migration from ServiceContext to Settings is documented, and it is not the only one, so code copied from a blog post written last year frequently will not even import. The library is split across a core plus a large number of separately versioned integration packages, which makes dependency resolution a normal part of your week and pinning mandatory rather than tidy. The high-level abstractions that make the first demo fast are precisely the ones you fight later: when retrieval returns the wrong thing, the only useful question is which chunks came back and why they scored highly, and the answer sits several layers below the interface you were sold. And the gravity is commercial. The smoothest paths run through the hosted service, so a free framework becomes a metered line in the client's bill unless you notice the drift and choose otherwise deliberately.

Spend an hour on it if you are looking at a folder of the client's genuinely messy documents and need something queryable this week. Skip it if the data is already clean text sitting in a database, where LlamaIndex is a large dependency to justify against a few dozen lines of chunking and a pgvector query you can read in one screen. Check the version number on every tutorial before you trust a line of it.

Write the ingestion yourself against pgvector when the corpus is uniform. Use LlamaParse alone when parsing is the only genuinely hard part.