Local vector store, retrieval, and grounded answers

Source text is chunked, embedded, and persisted in a local IndexedDB vector store. Queries retrieve the closest chunks and assemble a grounded answer from the evidence.

Global settings

1. Index source material

Add pasted text or drag in text-based files. Sources are chunked, embedded, and kept locally in IndexedDB so you can manage full documents and their stored chunks in one place.

Documents

0

Stored chunks

0

Indexed characters

0

Embedding model

all-MiniLM L6 v2

Sentence transformer · Not downloaded

Browser Storage

Database: IndexedDB

Capacity: Unavailable

Total used: Usage unavailable

Database data: Unavailable

Model cache: Unavailable

Storage policy: Unavailable

Embedding settings

Static word embeddings average fixed word vectors; sentence transformers encode the whole chunk.

Local embeddings

A sentence-transformer model that embeds the whole chunk for similarity search and retrieval.

Xenova/all-MiniLM-L6-v2

Category

Sentence transformer

Cache status

Not downloaded

Download an embedding model before indexing with it.

Chunking settings

Choose how source text is split before embeddings are stored.

Strategy

Store cap

500 chunks maximum

Indexing stops before this limit to keep browser memory, retrieval, and vector-map projection responsive.

Strategy

Keeps nearby sentences together until the chunk reaches a character budget.

Chunk budget

Set how much sentence-level context fits into each stored chunk before the next chunk begins.

Sentence-aware mode favors readable sentence groupings, word mode groups tokens by count, and recursive mode gives you tighter control over chunk length and overlap.

Upload files

Drop files here or add them through the picker to queue plain text, Markdown, JSON, CSV, HTML, XML, YAML, or code files with readable text.

Drop files here to queue source material.

Manual Entry

Type or paste one-off source material here when you want to create a document by hand. Uploaded files stay in the queue and index from there.

Document management

Stored documents keep the original source text so you can reindex with the current chunking settings or remove a document from local storage.

Indexed documents will appear here after you store source material.

Stored chunks

Each row below is persisted in IndexedDB. Removal updates the local vector store immediately.

No chunks are stored yet.

2. Ask the indexed corpus

The selected embedding model embeds the question, runs similarity search, and returns the closest chunks. A downloaded browser LLM then generates the grounded answer from those retrieved passages.

Selected LLM

Qwen3 0.6B ONNX

Cache status

Not downloaded

Question

Run retrieval against the indexed chunks, then generate an answer with the selected local LLM.

LLM settings

Select a small instruct model and download it with Transformers.js before running grounded answers.

Local generation

A Qwen3 browser model using ONNX weights with WebGPU and q4f16 quantization.

onnx-community/Qwen3-0.6B-ONNX

Category

Instruction model

Cache status

Not downloaded

Download a local LLM before running grounded answers.

Grounded answer

Download an LLM, then run a grounded answer from the top-ranked chunks.

Retrieved matches

Retrieved chunks will appear here after a query.

Semantic map

3. Explore the vector space

UMAP reduces each stored embedding into either two or three dimensions, and the force graph connects the nearest semantic neighbors. Retrieved matches are highlighted so you can compare retrieval results against the broader structure of the corpus.

Vectorized chunks

0

Vector dimensions

0

Retrieved highlights

0

UMAP settings

Adjust how the high-dimensional embeddings are projected into two or three dimensions.

Projection

Projection Basis

Choose the output space and the distance function UMAP uses before projection.

Structure

Control how local neighborhoods and cluster spacing are preserved in the reduced map.

Optimization

Tune repeatability and how much compute UMAP spends refining the layout.

Projection basis chooses the target space, structure shapes how clusters form, and optimization settings trade off repeatability and compute time.

Graph settings

Tune the displayed neighborhood graph that overlays the projected vector space.

Rendering

Connectivity

Control which semantic neighbors become graph edges.

Link Appearance

Adjust how prominent semantic links look once they are included.

Node Layout

Tune node size and on-screen spacing without changing the underlying vector relationships.

Connectivity controls decide which links appear, link appearance controls their emphasis, and node layout controls the graph's on-screen spacing.

Vector map

UMAP reduces each stored embedding to the selected 2D or 3D space, and the graph links the closest chunks by cosine similarity.

Orbit drag rotates and scroll zooms.

Index source material to generate embeddings before rendering the vector map.

Each dark node is a stored chunk. Retrieved matches appear in green, focused nodes turn blue, and the most recent query appears as an amber node linked to those matches. Nearby nodes usually indicate semantic similarity, but the on-screen distance is only approximate because UMAP, projection scaling, and spacing adjustments affect the layout.