DOCUMENTATION

AI

Where the assistant runs, what it is sent, and what it is not allowed to decide.

The agent explains readings and answers questions about the console. Verdicts come from the rule set alone: no model output enters a decision, and the assistant cannot create, change or close an investigation. On the Ledger Garden screen (Desk) a walked path, with its full addresses, hashes and evidence, can be handed to the agent as one question that asks it to use only those facts.

Providers

ProviderRunsCost to youKey
NOSHX CoreOn this deviceFreeNot needed
OllamaOn this deviceFreeNot needed
LM StudioOn this deviceFreeNot needed
llama.cppOn this deviceFreeNot needed
JanOn this deviceFreeNot needed
vLLMOn this deviceFreeNot needed
Claude (Anthropic)HostedProvider pricingRequired
OpenAIHostedProvider pricingRequired
GroqHostedFreeRequired
OpenRouterHostedProvider pricingRequired
DeepSeekHostedProvider pricingRequired
MistralHostedProvider pricingRequired

A custom OpenAI-compatible endpoint can also be configured. If you configure a hosted AI provider, your prompts are processed by that provider under its terms.

NOSHX

NOSHX is the agent on the NOSHX screen. It answers a question by reading the live ledger through the app's own read-only tools (issuer authority, order-book depth, settlements, control surfaces, provenance, pools, issuance, the address check, claims, NFT rights and ledger status), then explains what it read. It cannot sign, submit or move anything. Each tool carries the same plan requirement as the screen it mirrors, and address checks on the Free plan count against the same monthly allowance. Its reasoning is the chosen model's; its figures come from the ledger. A model that cannot call tools answers from the state it is given, and says so.

NOSHX also answers questions about NOSHASHI itself: its screens, features, plans and prices, and what these documentation pages, the Learn course, pricing, trust and legal pages say. Those pages are indexed inside the app (keyword search, no extra model and no network), the passages that match a question are given to the model with their addresses, and NOSHX can search them again as a tool. With no model running, the support desk answers from the same pages.

NOSHX Core

By default NOSHX runs on NOSHX Core, NOSHASHI's own engine, which uses no language model. It reads the question for what NOSHASHI can act on (addresses, transaction hashes, currency codes) and what is being asked about them, runs the matching read-only readers, and writes the answer from their reports with the findings the screens show. Product and compliance questions are answered with the most relevant sentences from these pages, with their sources, and questions about what to use name the screens that fit and the plan each needs. Core runs inside the app: nothing you ask is sent anywhere, only the ledger reads go to the public XRPL servers.

The support console on noshashi.app is NOSHX Core too, running in the visitor's browser. It answers from the website's support answers and these pages, reads an address, NFT id or transaction from the live ledger when a question names one, and hands anything it cannot answer to the contact form. A visitor has no plan, so paid readers say which plan they need; address checks count against the same free allowance of 10 a month, kept in that browser.

A language model is optional. A model that cannot call tools is given Core's readings to answer from; with failover on, a model that fails falls back to Core. Local requests use a fixed 8K context. Deep reasoning, off by default, lets a model think before answering; a model without a thinking mode is asked again without it.

Training NOSHX

NOSHX's own language model is trained from NOSHASHI's own content with the free kit in scripts/noshx-model. The training data is the in-app help, the Learn course's knowledge check and word list, the plan catalogue, the screen reference and every section of these pages. A Google Colab notebook fine-tunes IBM's Granite 4.0 1B (Apache 2.0 licence) on a free T4 GPU and exports a 4-bit file of about 1 GB, which answers quickly on a laptop's CPU and leaves most of an 8 GB laptop's memory free. Microsoft's Phi-4-mini-instruct (MIT licence, about 2.5 GB) is an option in the notebook when better writing matters more than speed. Add the model to Ollama with ollama create noshx -f Modelfile and choose it in the runtime panel. The model learns NOSHASHI's facts and voice; live ledger facts still come from NOSHX Core's readers at the time of the question.

The runtime can be switched at any time between a local model and a hosted one; each provider remembers its own endpoint and model. With failover on, a runtime that cannot be reached is replaced for that answer by the last one that worked, and the answer says which.

Support tickets

Signed-in customers open a ticket from NOSHX › TICKETS › NEW TICKET, with a subject, a topic, a priority and a description, and optionally the app version and platform. Each ticket gets a reference such as NSH-1042. NOSHASHI support replies in the same thread; the customer is emailed about a reply, and support is emailed about a new ticket or a customer's reply. A ticket reads WAITING ON SUPPORT until support answers, then SUPPORT REPLIED; either side can mark it RESOLVED, and a customer's reply reopens it.

Tickets are held in NOSHASHI's database with row level security: a customer can read only their own tickets, and only support staff can read every ticket. Messages cannot be edited or deleted by anyone. Every write goes through a server function that checks who is writing, and a ticket or reply containing an XRPL secret seed is refused in the app before it is sent. At most 10 tickets a day can be opened from one account.

Where API keys go

A provider key is kept in the operating system keyring. In the desktop app, model requests are made by the app itself rather than by its window: the key goes from the keyring to the provider's own host over TLS, and the window can store, check or clear a key but never read one back. A key for a named provider is only ever sent to that provider's API host, a custom endpoint must use HTTPS, and local runtimes never receive a key.

Secrets never reach a model

A message containing a valid XRPL seed, recognised by its checksum, is refused before any request is made.

Oversight

  • A policy submitted by one member must be activated by another; the server refuses self-activation.
  • Policy exceptions are requested by one member and decided by another, and every decision is written to an append-only audit log.
  • Investigations are opened, prioritised and closed by people. The AI assistant can be asked about a case but cannot create, change or close one.
  • Every call to an AI model is recorded with digests and sizes, so what was sent and when can be shown without keeping a second copy of the text.

Generated from src/lib/agent/providers.ts · src/lib/agent/secrets.ts · src/lib/trust/boundary.json