The search engine that thinks before it searches.
slorg drafts an answer and a knowledge graph from an LLM first, uses that graph to extract search keywords, queries the open web through SearxNG, then scores each result for relevance before it shows you anything. Same six steps every time. No magic.
Six fixed steps. No agentic loop.
slorg is not an agent. It does not decide whether to search; it always searches. It does not pick tools at runtime; the pipeline is the same on every query. What it does instead: spend its first round-trip on the LLM, not on Google, so that the keywords sent to the web are derived from a structured plan rather than from the raw user prompt.
Research teams who want their search legible
You can read the knowledge graph, see the extracted keywords, and inspect the per-result relevance score. The pipeline is six functions, not a black box.
People building RAG over the open web
slorg returns ranked, scored, content-extracted results as a JSON object.
The results[] array is shaped for ingestion into a vector store
or as in-context grounding for a downstream LLM call.
Perplexity users who want more control
Self-hosted. Bring your own OpenAI key (or any OpenAI-compatible endpoint
via OPENAI_BASE_URL). Pick your engines. Pick your model.
Internal-doc searchers after a model swap
By default slorg targets the open web. The relevance-scoring layer is source-agnostic; the README documents the public web backends, and the architecture page is where to look before pointing it at private indexes.
Install, set a key, ask.
npm install -g lorg
export OPENAI_API_KEY=sk-your-key-here
lorg "What causes the northern lights?"Or as a Node library:
import LorgSearch from 'lorg';
const lorg = new LorgSearch(process.env.OPENAI_API_KEY);
const results = await lorg.search('What is quantum computing?', {
model: 'gpt-4o-mini',
maxResults: 5
});
console.log(results.answer); // draft answer (step 1)
console.log(results.knowledgeGraph); // entities + relationships (step 2)
console.log(results.keywords); // extracted search terms (step 3)
console.log(results.results); // ranked web results with score (step 6)
console.log(results.tokenCount); // OpenAI tokens used
Or as a REST service: lorg server -p 3000 exposes POST /search.
Full API reference at
docs.skelfresearch.com/slorg.
Five env vars. That is the whole surface.
| Variable | Default | Purpose |
|---|---|---|
| OPENAI_API_KEY | — | Required. Authenticates the LLM calls in steps 1, 3, and 6. |
| OPENAI_BASE_URL | OpenAI | Point at any OpenAI-compatible endpoint (local llama.cpp, Together, Groq, etc). |
| LORG_DEFAULT_MODEL | gpt-4o-mini | One model for all three LLM steps. No per-step override yet. |
| LORG_SEARCH_ENGINES | google,bing,yahoo,duckduckgo | Comma list passed to SearxNG. |
| LORG_SEARCH_LIMIT | 20 | Max URLs returned by SearxNG before content-fetch and scoring. |
Source: project README. If you need a knob that isn't on this list, it isn't wired up — open an issue.
-
Source-aware retrieval: knowing what your engine knows
Most search engines hide their backends behind a single answer. slorg shows you which engines returned which URLs, what content was actually extracted, and what the LLM scored each one — because retrieval without source awareness is just confident guessing.
-
Planning the search: where the latency budget really goes
A breakdown of the three LLM calls and N content fetches that make up a slorg query, what each one costs, and which knobs actually move the needle.
-
Most agentic search throws queries at Google and prays
Why the dominant pattern in LLM-mediated search is functionally a wrapper around a single search call — and what changes when you make the plan the first artifact, not the last.
Narrow, honest comparisons against tools that overlap on one axis or another.
-
slorg vs. Perplexity
Both produce an answer plus ranked sources. One is a hosted product; the other is six functions you can read.
-
slorg vs. SearxNG + your own LLM glue
If you already run SearxNG and a model, what slorg adds is the plan: a graph, extracted keywords, and a scoring pass — in that order.