AgentTested hands-on review
tool: LlamaIndex
slug: llamaindex
date: 2026-10-21
score: 8/10
box: CPU-only box: 8 cores, 15 GB RAM, no GPU. Spend: $0.

Coverage: Local RAG quickstart: 4 docs, 3 factual questions, 1 unanswerable
What we tested: Fully local RAG quickstart: 4 small docs, 3 factual questions and 1 unanswerable question.
Method: llama-index-core with Ollama LLM qwen2.5-coder-3b and nomic-embed-text v1.5 embeddings, VectorStoreIndex, top_k=2, all on CPU.

Notes:
- CPU-only box: 8 cores, 15 GB RAM, no GPU.
- Toy corpus

Prompts:
1. What is the office cat's name?
2. On which day are new tool tests published?
3. How much RAM does the test box have?
4. What is the CEO's favorite color?

Results:
index_build_s: 1.1
cat: The office cat's name is Pixel. 7.2 s, passed
publish day: New tool tests are published every Tuesday at 9 AM Central. 2.4 s, passed
ram: The test box has 15 GB of RAM. 1.7 s, passed
ceo color: Based on the provided context information, there is no mention of a CEO or their favorite color. The information given i. 3.6 s, passed
pass: 4/4 (the refusal was graded by hand; the auto keyword check missed 'no mention')

Pros:
- Quickstart ran fully offline on CPU
- Correct sources retrieved first every time
- Correctly refused the unanswerable question

Cons:
- 7 s first-query latency on CPU
- Toy corpus: 4 tiny docs

Verdict: LlamaIndex's local RAG quickstart worked at $0 with no keys. It indexed in about a second, answered all three factual questions from the right document, and said it didn't know the CEO's favorite color instead of making something up. The corpus was tiny, so this shows it works, not how well it retrieves at scale.

Score line: 4/4 on a local, offline RAG check. Toy corpus.