StackMap
Subscribe

LlamaIndex vs searchbox

Data framework for connecting custom data sources to LLMs — ingestion, indexing, retrieval. — versus — Airgapped closed-corpus QA testbed: a local Qwen agent in a Pi harness explores a .zip dataroom with grep/embeddings/rerankers under a token budget — a bed to study search as test-time compute.

The curated verdict

Both answer questions over a private corpus, but by opposite paradigms. LlamaIndex is a production RAG framework: build a structured index, then query it. searchbox is an airgapped research harness where an agent explores the raw corpus with grep/embed/rerank tools under a budget — no prebuilt index. Use LlamaIndex to ship RAG, searchbox to study agentic retrieval.

LlamaIndexsearchbox
Stars51k49
Forks7.8k6
LanguagePythonPython
LicenseMITMIT
Last activityyesterday21 days ago
Topicsrag, agentslocal, rag, agents
Curated connections83

LlamaIndex — the curator's take

Data framework for connecting custom data sources to LLMs — ingestion, indexing, retrieval.

searchbox — the curator's take

A research testbed, not a product: it locks a local Qwen3.6 agent in an airgapped box with a .zip 'dataroom' and makes it answer purely from what's inside, using grep, embeddings (jina-v5) and a cross-encoder reranker (jina-v3) as tools under a turn budget. It exists to ablate one question — is search test-time compute? — so every axis (which tools, local-weights vs Jina-API retrieval, budget, base LLM) is an env knob and each run emits per-turn token/tool telemetry. Reach for it to study or benchmark agentic retrieval over your own corpus, self-hosted and offline. NOT a drop-in RAG library or production QA service: no persistent index (it embeds lazily per job), one job at a time on a single model slot, and you bring your own OpenAI-compatible LLM endpoint (llama.cpp/ollama) plus ideally a GPU.