v0.5.2

Vectorless, reasoning-based RAG in JavaScript. Builds a table-of-contents tree from a document and lets an LLM navigate it with tools, instead of chunking and embedding.

bun add vectorless-js

Tree quality

buildIndexAuto now routes on why a tree came out flat, because the two causes cost very differently.

Some PDFs give the parser no font-size hierarchy, so every heading parses at one level. The titles and line numbers are already correct and only the nesting is missing. assignHeadingLevels sends the model the headings alone, batched, and nests them; tree assembly stays deterministic. Measured on three General Mills 10-Ks: 318 flat nodes become 53 top-level at depth 5, node count preserved, on a few hundred tokens per call instead of the ~133k a full reconstruction would have sent. That prompt did not fit a 128k context window, so those documents previously could not be indexed at all.

A document with no headings still goes through reconstructTree, which infers the outline from the body.

Depths come back as number:depth pairs keyed by heading number, so a heading the model skips cannot shift the rest, and an unusable answer degrades to the flat tree rather than a wrong one.

Branding

The agent's system prompt introduced itself to the model as PageIndex on every ask(). It now says vectorless-js. Attribution for the port stays in LICENSE, README, and source comments.

Packaging

solid-js moved to devDependencies; it was shipping SolidJS to every consumer. gpt-tokenizer is the only runtime dependency. LICENSE now carries Vectify AI's notice alongside ours, as MIT requires for the ported tree builder.

Also

Quick-start script guards its argv, so it typechecks under bun init's default noUncheckedIndexedAccess and prints usage instead of crashing. New runnable example in examples/quick-start/.

Full package: https://www.npmjs.com/package/vectorless-js