Loading…

Dashboard

System overview and metrics

0
Documents
0
Chunks
0
Index terms
0
Queries
0 ms
Avg query time
–
Storage used
–
Embedding
–
TurboQuant v2

System Health

Recent questions

About Abida

Artificial Business Intelligence Database Applications

A private knowledge base with a RAG-style question answering interface. The search engine is written in Rust and runs as WebAssembly in a Web Worker: BM25F keyword index, hashed n-gram embeddings compressed with TurboQuant v2, hybrid ranking (RRF + MMR) and extractive answers with citations. Everything stays in this browser.

니 애비다 (I am Your Father) · 앱이다 (It is an App)

Chat

Ask questions about your knowledge base — answers quote their sources

Knowledge Base

Add and manage documents

+

Drop files here to learn (or click)

TXT · MD · CSV · JSON · HTML · PDF · Abida desktop JSONL

Write Knowledge

Learn from Web

Free places to learn from — click one:

Wikipedia, Wikisource and Wikibooks are read through their open APIs; Project Gutenberg books come from the archive.org copy. Other sites may block reading from another page.

Keep studying — until you stop

Pick a list; Abida reads and learns one item after another and shows the queue. Stop whenever you like — everything learned so far stays.

    Stored Documents

    Brain · 두뇌 동기화

    A small language model in this browser learns your documents — LoRA fine-tuning on WebGPU

    –
    Compute
    0
    Training steps
    –
    Loss on your documents
    –
    Tokens / second

    Learning curve

    Ask the Brain directly

    The Brain writes by continuing your text. After a sync it knows the phrasing of your documents; it is small (a few million parameters), so treat it as an experiment — the Chat answers with citations stay the reliable ones.

    Settings

    Every setting here changes what the engine does

    Chunking & index

    Documents are split at headings, paragraphs and sentences into chunks; neighbouring chunks overlap a little. BM25F weighs matches in the title and section path higher than in the body.

    Ranking & answers

    Synonyms & stop words

    TurboQuant v2

    Chunk vectors are stored as 4-bit codes after a seeded Hadamard rotation, with a scale/offset per 32 dimensions (f16). Search multiplies the f32 query with the 4-bit codes directly (SIMD), nothing is decoded. Vectors of the desktop app (TurboQuant v1) remain readable.

    Vectors
    –
    Bytes / vector
    –
    vs f32
    –
    Vector store
    –
    Postings
    –