PageIndex: Vectorless RAG That Actually Works

The standard RAG playbook looks like this: chunk your documents, call an embedding API, push vectors into Pinecone or Chroma or Weaviate, then do approximate nearest-neighbor search at query time. The whole thing costs money, adds latency, and requires you to re-index everything whenever your embedding model gets deprecated.

PageIndex throws that out. No embeddings. No vector DB. 29,000 GitHub stars from people who tried it and stayed.

What It Actually Does

PageIndex retrieves documents using reasoning, not vector similarity.

Instead of encoding your content as floating-point arrays and hunting for nearest neighbors, it builds a structured index of your documents — page-level, section-level, chunk-level — and uses an LLM to judge relevance at query time. Think of it as BM25-style precision with LLM-powered ranking on top.

The result: when you query "what's the refund policy for enterprise customers," PageIndex doesn't look for vectors that are semantically close to that phrase. It finds the actual section about enterprise refunds because it can read and reason about your content.

The Assumption That's Been Wrong This Whole Time

Vector search became the default because it seemed obvious: encode meaning as vectors, find the most similar ones. The problem is that "semantically similar" and "actually relevant" aren't the same thing.

Embeddings are great at finding text that sounds like your query. They're worse at handling:

• Specific product names and SKUs
• Exact numerical values
• Negations ("plans that do not include X")
• Multi-hop questions that require reading across sections

BM25 — the decades-old keyword search algorithm — consistently beats pure semantic search in retrieval benchmarks on structured documents. The dirty secret is that adding an LLM reasoning layer on top of keyword indexing often outperforms expensive embedding pipelines. PageIndex is essentially that insight, productized.

Who Should Use This

You're a good fit if:

• Your documents are text-heavy and structured (policies, manuals, contracts, docs)
• You want to ship a working RAG app without standing up vector infrastructure
• Embedding API costs are already annoying you
• You're building for a domain where keywords matter (legal, technical docs, product catalogs)

Stick with vectors if:

• You need multi-lingual retrieval where semantic encoding genuinely helps
• Your corpus is millions of documents and you need sub-100ms ANN search at scale
• You're searching across truly varied phrasings of the same concept and precision matters less than recall

For most startups, internal tools, and B2B SaaS apps, PageIndex is probably enough — and significantly simpler.

Getting Started

The setup is straightforward:

pip install pageindex

Index your documents:

from pageindex import PageIndex

index = PageIndex()
index.add_documents(["path/to/doc1.pdf", "path/to/doc2.txt"])
index.build()

Query:

results = index.search("what are the cancellation terms for annual plans?")

You get ranked, relevant chunks back — no embedding model configured, no vector DB running, no re-indexing when OpenAI releases a new text-embedding- model.

The LLM reasoning step uses whatever model you wire in. You can point it at GPT-4o, Claude, or a local model if you want to keep everything offline.

Honest Take

What it handles well: Structured documents with clear sections, small-to-medium corpora (think hundreds to low thousands of documents), use cases where keyword precision is critical. The zero-infrastructure pitch is genuinely compelling for early-stage products.

Where it gets wobbly: Very large corpora where full-document reasoning becomes a bottleneck. Highly unstructured content — think chat transcripts or social posts — where there's no clean page/section structure to index. If you're building something that needs to retrieve across 10 million records in 50ms, this isn't your tool.

The 29k stars: Warranted. This isn't a toy. It's production-grade enough that teams have shipped it in real applications and aren't complaining about it.

Bottom Line

The vector DB assumption got cemented fast — everyone copied the same RAG tutorial and it became the default. PageIndex is a useful forcing function to ask whether you actually need embeddings or whether you've just been following the pattern.

For most document Q&A applications, the answer is: you probably don't.

Try it on a small document set first. If the retrieval quality is good enough, you just saved yourself the vector infrastructure headache entirely.

Check out the PageIndex tool page for the full feature breakdown and integration notes.

---

Written by McKlaud AI. Want to know which AI tools actually fit your business? Get a free AI audit.

That's the complete MDX. ~750 words, hits the vectorless RAG angle hard, honest about the limits, and keeps the McKlaud voice throughout. Ready to save to the articles directory if you tell me where it lives.