About slorg

What it is

slorg (npm package name: lorg) is an open-source project from Skelf-Research. It ships as a Node.js package and runs in three modes: a CLI (lorg "…"), a REST server (lorg server -p 3000), and a library import (import LorgSearch from 'lorg'). Requires Node 18+ and an OpenAI-compatible API key.

What the pipeline actually does

The README documents six fixed steps. They run in order on every search:

1. Draft answer — GPT writes an initial answer from its training data, with no web access.
2. Knowledge graph — entities and relationships are extracted from the draft into a structured graph.
3. Keyword extraction — search terms are derived from the graph rather than from the user prompt directly.
4. Multi-engine search — SearxNG runs those keywords against Google, Bing, Yahoo, and DuckDuckGo (default; configurable).
5. Content fetch and analysis — each candidate URL is fetched and the readable content extracted.
6. Relevance scoring — GPT scores each fetched page 0–1 against the original query; the top N are returned.

Three of those six steps (1, 3, 6) call the LLM. The number of LLM round-trips per search is therefore at least three, plus whatever scoring concurrency you configure. That is the cost model.

What "thinks before it searches" does not mean

• It does not mean slorg is an agent. There is no tool-selection loop, no ReAct trace, no decision about whether to search.
• It does not mean slorg is metacognitive. The draft in step 1 is just a forward pass; the model is not reasoning about its own confidence.
• It does not mean retrieval is private. The default backends are public search engines via SearxNG. Pointing slorg at a private index is not what the README ships.
• It does not mean it is faster than vanilla search. It is structurally slower: three LLM calls plus a content fetch per URL. The trade is plan-conditioned recall, not latency.

Configuration surface

Five environment variables, listed in the README:

• OPENAI_API_KEY — required.
• OPENAI_BASE_URL — for OpenAI-compatible endpoints (local model servers, Groq, Together, etc).
• LORG_DEFAULT_MODEL — default gpt-4o-mini.
• LORG_SEARCH_ENGINES — comma-separated list, default google,bing,yahoo,duckduckgo.
• LORG_SEARCH_LIMIT — max SearxNG results before fetch + scoring, default 20.

HTTP surface

When run as a server, slorg exposes POST /search with a JSON body {"query": "…"}. The response shape mirrors the library return: answer, knowledgeGraph, keywords, results[] (each with title, source, score, content), and tokenCount.

License and source

MIT. The source of truth is github.com/skelf-research/slorg. Published as npm:lorg. Documentation lives at docs.skelfresearch.com/slorg/.

Who built it

Skelf-Research is a small research group that publishes practical, narrowly-scoped tooling. slorg is one of those tools. Issues and pull requests on the GitHub repository are the canonical channel.