nanograph 1.3.0

1.3.0 reshapes the query language around one principle: the agent is the user. A query is a tool. The .gq file is the toolbelt. Workflows are sequences of tool calls the agent composes — not multi-stage queries the human composes for them.

New mutation syntax follows directly from that. put upserts by @key in a single statement, so the agent stops writing insert-then-update dances. where { ... } blocks express conjunctive CAS guards — newline-separated atoms, including is null for first-claim — so a claim only goes through if every predicate still holds. Lance OCC plus the row gate guarantee exactly-one-winner under concurrent claims. Every mutation response wraps the affected rows in a matched_nodes envelope, so the agent verifies what it touched without a second query.

On the embedding side, LM Studio joins OpenAI and Gemini as a third provider. Set NANOGRAPH_EMBED_PROVIDER=lmstudio, point at your local LM Studio server, and the same @embed(...) and nearest(...) flow runs entirely on-device. Pairs naturally with on-device LLMs for a fully-local agentic stack.

The runtime got noticeably faster while we were at it. Coordinated bump to Lance 6, Arrow 58, DataFusion 53 pulls in SIMD-accelerated distance kernels, a RaBitQ 4-bit LUT (16× speedup on ARM), and a segmented inverted index for full-text search. Per-runtime hot-path caches eliminate the most common waste in agent loops: query-text embeddings now cache across calls, put upserts skip the dedupe rescan when the source side is already unique, and compiled query plans are cached across executions.

The full design rationale for the language direction lives in ql-canon.md at the repo root.

What shipped

Agent-toolbelt mutations

Mutations safe to chain in agent loops — atomic, idempotent, with a receipt.

put — upsert by @key in a single statement; replaces the "try insert, fall back to update" dance
where { ... } — conjunctive CAS guards (newline-separated atoms), with is null for first-claim
Exactly-one-winner under concurrent claims, guaranteed by Lance OCC plus the row gate
matched_nodes envelope on every mutation response — the agent verifies what changed without a second query

Agent toolbelt design→

LM Studio embedding provider

Third local-embedding option, alongside OpenAI and Gemini. Completes the on-device story.

Set NANOGRAPH_EMBED_PROVIDER=lmstudio; defaults to http://localhost:1234/v1
Same @embed(...) and nearest(...) flow as the other providers — no code change
Runs entirely on-device; pairs naturally with on-device LLMs for a fully-local agentic stack
Text-only, like OpenAI. Model is required (no safe default — must match what's loaded in LM Studio)

Embeddings guide→

Faster, modernized engine

Coordinated bump to Lance 6, Arrow 58, DataFusion 53 — plus per-runtime hot-path caches.

SIMD-accelerated u8/bf16/f16/f64 distance kernels for vector search
RaBitQ 4-bit LUT (16× speedup on ARM), segmented inverted index for FTS, eager I/O scheduler
Query-text embeddings cached across calls — no re-embed loops when an agent repeats nearest($x, $q)
put upserts use SourceDedupeBehavior::FirstSeen — skip the dedupe rescan when the source side is unique
Compiled query plans cached per (schema, query) pair — re-runs skip parse + typecheck + lower

Upgrading

brew upgrade nanograph or grab the binary from the release. Rust crates and the npm package are at 1.3.0. On-disk format is compatible — graphs created on the previous binary open cleanly on 1.3.0.

Star the repo if you want to follow along.