Quesarby MLAI

Docs

Start with the source. Stay on the site.

Sidebar, search, and articles are the same corpus. Setup commands are copied from READMEs. A successful gate is evidence for that checkout — not for a hosted product.

Security & trust

Fail closed. Tag the figure.

Sessions, evaluation gates, and the difference between measured, reported, and target.

1 min read

Security & trust

Evidence before promises.

Keep documented behavior, reported testing, measured results and targets separate.

1 min read

Quesar developer platform

Build private, traceable AI workflows on the ABI runtime: retrieval provenance and hash-chained memory through WDBX, exposed over a local CLI and an MCP server. Policy-gated agents and an evaluation mesh are planned, not shipped.

Traceable retrieval

Partial

Index records with source metadata and inspect weighted backtrace paths and retrieval score components; snapshots and write-ahead recovery give a restore point.

Agent policy gates

Planned

Planned: bind tools to explicit permissions, approval thresholds, and review roles before execution reaches production data. Today the MCP server only fails closed on unknown tools.

Evaluation mesh

Planned

Planned: regression suites for retrieval faithfulness, prompt-injection resilience, latency, and operator review burden, run as a release gate.

Private runtime

Partial

The ABI CLI and MCP server run locally on operator-owned machines. Packaging for cloud, VPC, on-premise, and offline-first deployments is the design, not a shipped installer.

Start

ABI runtime

ABI is a Rust framework for local AI orchestration, semantic vector storage, and GPU capability reporting. Build the CLI and MCP server with the repository's pinned toolchain and the ./tools/cargo.sh wrapper.

donaldfilimon/abi
# Validate the Rust workspace
./tools/check.sh
# Build the CLI and MCP server
./tools/cargo.sh build -p abi-cli -p abi-mcp

# Inspect capabilities and terminal surfaces
./target/debug/abi backends
./target/debug/abi dashboard --pane system --once --json
./target/debug/abi agent tui
Runtime
Rust workspace: AI, WDBX, GPU, MCP, terminal
Configuration
Crate-specific Cargo features
Capability inspection
abi backends · abi wdbx gpu info

  • abi-ai

    Profile routing, completion, and governance helpers.

  • abi-sea

    Evidence selection, scoring, and learning loop.

  • abi-wdbx

    Durable memory and retrieval from the sibling Rust substrate.

  • abi-gpu

    Backend reporting, optional Metal DOT kernels, and CPU fallback.

  • abi-mcp

    JSON-RPC tool handlers and stdio server.

  • abi-cli

    Commands, agent REPL, and diagnostics dashboard.

Design decisions

Inspectable capabilities

Report the selected backend and whether acceleration is active. Capability detection alone is not evidence of accelerated execution.

Bounded evidence

SEA selects evidence within record, token, cluster, and prompt-byte budgets. Retrieval is distinct from model training.

Explicit runtime boundaries

Local completion is deterministic persona-template generation. Live HTTP completion requires an explicitly configured provider.

Architecture

How a profile is selected

An explicit leading persona address takes precedence. Otherwise, token-prefix signals adjust a prior and the runtime normalizes the resulting weights before selecting the largest. The distribution describes routing preference, not model quality, authorization, or a parallel-execution strategy.

Explicit address

A leading Abbey, Aviva, or Abi name selects that profile; mentioning a name later in prose does not.

Token-prefix signals

Without an explicit address, keyword stems at the start of whitespace-separated tokens adjust an Abbey-favoring prior.

Normalized selection

The largest normalized weight selects the primary profile. A routing share is not a calibrated confidence in correctness.

Abbey's voice

Abbey is the profile you hear most in explanation and review, so her voice sets the tone for the framework: Intelligence Without Limits, gated by policy. She says what she knows and what she doesn't, and will not claim unlimited or AGI capability, unverified benchmarks, or Quesar as a bot feature.

“Care first. Clarity always. Competence throughout.”

Care first

Read the person's goal and state before reaching for the answer; meet them where they are, never condescending.

Clarity always

Explain the why, not just the what; teach rather than dictate, and keep jargon in service of understanding.

Competence throughout

Broad technical range, paired with the honesty to name uncertainty and defer to review instead of bluffing.

Read the paper: policy-locked tool use

Architecture

WDBX retrieval

WDBX is the Weighted Directed Backtrace eXecution store. It keeps context as weighted paths so retrieval can be inspected, not just ranked: key-value, vector, and block surfaces, with snapshot persistence guarded by SHA-256 integrity checks.

Weighted backtrace paths

Partial

Hybrid retrieval exposes its score components. The score still collapses several axes; separate evidence-weighted ranking is planned.

HNSW vector search

Current

Cosine nearest-neighbor through the active Rust substrate's layered HNSW index (M=16, EF_CONSTRUCTION=40, EF_SEARCH=32).

Durable snapshots

Current

Snapshots with CRC-framed WAL recovery and checkpoints. Snapshot-chain and strict content verification are separate checks.

Switchable persistence

Current

The local store defaults to $HOME/.abi/wdbx; ABI_WDBX_PERSIST=0 disables persistence for evaluation runs.

Read the paper: WDBX weighted-backtrace store

Architecture

WDBX V2 documentation snapshot

A frozen June 2026 snapshot of the earlier single-crate Abbey WDBX project (Rust 2024, stable toolchain). It is retained for reference and is not the guide to the current abi-extracted substrate; claims in it require revalidation there.

Read the release note: WDBX V2

Architecture

MCP server

The abi-mcp server speaks JSON-RPC 2.0 over stdio. Its optional loopback HTTP compatibility listener serves persistent MCP 2024-11-05 HTTP+SSE sessions; it is not Streamable HTTP. Tool names and transport facts are from the ABI README.

Transport
JSON-RPC 2.0 over stdio
Request cap
64 KB
Loopback HTTP listener
127.0.0.1:8080 (optional)
Port override
ABI_MCP_HTTP_PORT
Bearer auth
ABI_MCP_HTTP_TOKEN (optional)
Session stream
GET /sse: persistent MCP 2024-11-05 HTTP+SSE session
Message endpoint
POST /message; without a sessionId, one-shot direct reply

Tools

  • ai_learn

    Learning entry point with local side effects; ABI_WDBX_PERSIST=0 disables persistence.

  • scheduler_info

    Compatibility alias for scheduler statistics.

  • ai_complete

    Local template completion through the routed persona profile; the requested model id is metadata, not proof it ran.

  • ai_run

    Run completion with local profile routing.

  • ai_train

    Train the selected local profile against WDBX.

  • wdbx_query

    Vector / block retrieval against the WDBX store with ordered results.

  • wdbx_stats

    Report local WDBX store statistics; not a cluster dashboard.

  • gpu_status

    Report GPU capability and backend, with deterministic CPU fallback.

  • scheduler_stats

    Report scheduler task counts.

  • connector_test

    Run local connector validation; does not prove live credentials work.

  • plugin_list

    Enumerate registered plugins and their target features.

  • plugin_run

    Invoke a registered plugin entry point.

Operations

Configuring this site

For quesar.cloud itself. Each missing value produces a clear "not configured" state, never a crash.

  1. 01

    Database

    Set DATABASE_URL for Postgres. Without it the app runs on in-memory PGLite, and data does not survive a restart or a serverless instance.

  2. 02

    Encryption key

    Set APP_ENCRYPTION_KEY (32 bytes, openssl rand -base64 32). Without it, sealed features refuse instead of storing plaintext.

  3. 03

    Server-only provider keys

    Set XAI_API_KEY, or the Cloudflare AI Gateway URL, token and id, plus LLM_PROVIDER. Never expose them to browser bundles.

  4. 04

    Administrators

    Set ADMIN_EMAILS. Only an allowlisted address with a linked Google or Apple account becomes an admin.

  5. 05

    Evaluation gates

    Run evaluation gates before allowing autonomous write actions or external tool calls.

Reference

What this site exposes

Protected surfaces require a Better Auth session and fail closed. There is no public hosted assistant API.

  • /api/auth/*

    Current

    Better Auth: Google, Apple and X sign-in, passkeys, email and password, session.

  • askPersona

    Current

    Server function behind the session: one persona reply through the model interface, rate-limited per user.

  • askDesk

    Current

    Server function behind the session: a desk answer from the catalog and, when configured, the model.

  • Console notes

    Current

    Per-user field notes on architecture nodes, scoped by user id.

  • GET /feed.xml

    Current

    RSS 2.0 for lab notes, research publications and releases. Public.

  • POST /api/telemetry, POST /api/csp-report

    Current

    Operational sinks: allowlisted anonymous page events (rate-limited, size-limited) and browser CSP violation reports (logged, not stored).

  • GET /api/cron/audits-expire

    Current

    Daily expiry of old chat audits. Requires Authorization: Bearer CRON_SECRET; without the secret it answers 503 and is off.

  • Consent, audits, admin review

    Current

    Consent-gated chat with sealed audits you can read and delete; admin review with a stated reason.

  • Workspace connectors, billing, inquiries

    Current

    Drive and SharePoint connectors, profile billing, and the public inquiry form with a rate limit.

Get started

Next

Architecture

The same layers, as a diagram.

Next

Developers

Live READMEs when GitHub answers.