-
-
Notifications
You must be signed in to change notification settings - Fork 3
FAQ
Is Nous a RAG system?
No. RAG (Retrieval-Augmented Generation) retrieves text chunks from a vector store and appends them to the prompt. Nous extracts typed relational structure from your content — concepts, the connections between them, and the evidence strength of each connection. The injected context block is compact, structured, and semantically precise. It is not a chunk of text.
Does it work offline?
Yes. Nous itself runs entirely locally — the graph is stored on your machine at ~/.local/share/nouse/field.sqlite. The daemon can use a local LLM (via Ollama) for extraction and enrichment. You only need an internet connection if you configure a cloud provider like Groq or Cerebras.
Does it work with any LLM?
Yes. Nous is provider-agnostic. It injects a context string — you handle the LLM call. Any model that accepts a system prompt works: OpenAI, Anthropic, Mistral, Groq, Cerebras, Ollama, LM Studio, or a raw HTTP endpoint.
Can it compete with NumPy in performance?
Nous is not a numerical computing library — it is a memory substrate for LLMs. It does not compete with NumPy any more than a database competes with NumPy. The comparison does not apply.
What database does Nous use?
SQLite WAL + an in-memory graph runtime. No external server required. The graph lives in a local file on disk and the daemon owns the write path.
Important: the daemon should own graph writes. If you are doing direct graph operations, stop the daemon first:
systemctl --user stop nouse-daemonHow much disk space does the graph use?
Depends on how much you feed it. A typical knowledge graph for a single software project or research domain uses 50–200 MB. SQLite WAL is compact enough for normal local use, and NightRun prunes weak edges nightly to keep the graph lean.
Can I have multiple graphs?
Not yet in the current version. Multiple graph support is planned. For now you can set a custom graph path:
brain = nouse.attach(path="/custom/path/to/field.sqlite")What happens when the daemon is not running?
brain.query() still works — it opens the graph in read-only mode. You just won't get new content extracted or Hebbian updates applied until the daemon is running again.
Does Nous store my data anywhere external?
No. The graph is local. If you configure a cloud LLM provider (OpenAI, Groq, etc.) for extraction, your content is sent to that provider's API during the extraction step — the same as any LLM call you would make manually. Nous does not add any additional telemetry or external storage.
The 96% result seems too good. What's the catch?
A few important caveats:
- The questions were generated from the same graph that was used to answer them. This tests whether the model can use the graph correctly — not whether the graph contains true knowledge.
- Scoring is keyword-based, not LLM-judged. Some Config A answers scored 0 due to API rate-limit errors, which inflated the gap slightly.
- The domain is Björn's own research graph — domain specificity is high, which is exactly where Nous helps most.
The result is directionally correct and reproducible. It is not a general claim that Nous makes every 8B model beat every 70B model on every task.
→ See Benchmark for full methodology.
Can I run the benchmark on my own domain?
Yes — that's encouraged. See Contributing for how to add your own question bank and submit results.
Who built this?
Nous is built by Björn Wikström at Base76 Research Lab.
What does "Nous" mean?
νοῦς (nous) is the ancient Greek word for mind, intellect, or reason. Aristotle used it to describe the faculty that apprehends first principles directly. It felt like the right name for a persistent domain mind.
Is this production-ready?
Not yet. Nous is research-grade software. The core API is stable enough to use, but the daemon, NightRun, and Ghost Q components are still maturing. Use it, break it, report what you find.