Alexandrina
drasticstatic · engine preview

Alexandrina.

AstralLogicEndpointeXpander:AutomatedNeuralDistributedResourceInferringNexusArchitecture

A small, careful knowledge engine for communities. Point it at your own materials, and members get answers in your own words, with the page they came from, plus helpdesks that turn requests into tracked work. When it doesn't know, it says so.

Alexandrina: a portrait woven from ember light
One engine, many libraries

Every community keeps its own shelf

Each community is a tenant: its own sources, voice, crisis text, desks and visibility rules. The engine never mixes shelves, and committee-only drafts never leave committee. That's enforced in code, with tests that try to leak them.

Sourced or silent

Answers cite their pages. No match means an honest “not in our materials.”

Safety before AI

Crisis, dosing and sourcing are caught by plain rules before any model runs.

An honest model chain

Claude when funded, then NVIDIA NIM. If all fail: “unavailable”, the closest pages and a ticket.

Private by design

No content logging. Ticket requesters are stored as a keyed hash, never an ID.

The pipeline

Eight steps, three honest exits

Every question walks the same path. The dashed gates are where it may stop on purpose: for safety, for lack of evidence, or because every model is down.

01
Questiontrimmed, length-checked
02 · gate
Safety guardcrisis → resources · dosing → refusal
03
Visibilitypublic surfaces search public shelves
04
Two searchesBM25 keywords + embeddings
05
Fusionreciprocal rank, k = 60
06 · gate
Confidencenone → “not in our materials”
07 · gate
Model chainClaude → NIM → … · 15 s each
08
Cited answeror honest unavailable + ticket
 tenants/your-community/tenant.yaml
id: your-community
name: Your Fellowship
persona: { name: A.L.E.X., file: persona.md }
sources:
  - { id: site,      type: markdown-dir, visibility: public }
  - { id: committee, type: markdown-dir, visibility: committee }
crisis:  { file: crisis.md }    # shown verbatim, no AI
refusal: { file: refusal.md }
retrieval: { topK: 6, minScore: 4 }
desks:
  - id: techcom
    command: techcom              # → /techcom
    prefix: TECH                  # → TECH-12
    githubRepo: your-org/techcom-desk
Configuration, not code

A new community is a folder

Persona, sources, crisis text, retrieval thresholds and helpdesks live in one YAML file beside the community's own Markdown and PDFs. Deployments snapshot the engine instead of forking its history, so private origins stay private.

700

lines of core engine

66

tests passing

0

invented answers, by design

A.L.E.X., Astral Logic Endpoint eXpander, PIR®'s Discord helper
Flagship tenant · PIR® Psychedelics In Recovery™

A.L.E.X. is live

The first deployment serves PIR®, a twelve-step fellowship, in Discord. /ask answers from PIR's own pages and literature, and the LitCom, TechCom, GitHub, PR and moderation desks open private, tracked tickets. An alexandrine is a twelve-syllable line, which suits a home that walks with Twelve Steps and Twelve Traditions, and A.L.E.X. has one of her own.

Roadmap

Where it's headed

Pilot build, October 2026. The engine source is private for now; this page is its public window.

  • Discord surface: slash commands, a chat panel, private ticket threads, archive and reopen
  • Helpdesks → GitHub: problems become Issues, ideas become Discussions
  • FAQ answer cache: instant replies that survive model outages
  • Citations that land on the sentence: text-fragment links
  • Web portals with password-free magic-link sign-in
  • Research lanes via Firecrawl, kept apart from community shelves
  • More tenants: O-R-G next, with its community's blessing