Qualia-DB | Online LLM Demo

Model

Starting WASM engine…

Select verified browser model

Q4_K_M browser presets are temporarily withheld: the current WebGPU quantized path can terminate on the first token. This Q8_0 image is the locally coherence-tested profile.

Or load a local P64 / GGUF file

The verified GGUF is compiled to canonical P64 once on-device and cached in OPFS. Local P64 / GGUF remains supported. Inference runs through Qualia's native CPU-WASM backend, with WebGPU used as an optional accelerator (qualia_core_db_bg.wasm).

Data Sources

QualiaDB Local

localhost:4242/query (SPARQL)

Wikidata SPARQL

query.wikidata.org/sparql

Inject Ontologies

QualiaDB vocab in system prompt

Mode

Chat Text→SPARQL JSON-LD Lookup

General conversation. The model answers freely, optionally grounded by SPARQL context.

System Prompt

Conversation

1) Open Setup (phones) · 2) Tap Load Model · 3) Chat fully on-device via CPU-WASM or WebGPU.

Shift+Enter = newline

Generated SPARQL

-- SPARQL will appear here after generation

JSON-LD Output

-- JSON-LD will appear here

Architecture: LLM ↔ SPARQL closed loop

The pipeline runs entirely client-side via Qualia's native inference stack (gguf_bridge → CPU-WASM or WebGPU decode). No third-party llama.cpp bindings. P64/GGUF only — not LiteRT-LM. On phones prefer SmolLM2-360M.

[ Human Input ]
       │
       ▼
┌──────────────────────┐
│ Local P64 / WASM LLM │ ◄── System prompt with QualiaDB ontologies
└──────────┬───────────┘
           │ Emits SPARQL query or JSON-LD
           ▼
┌──────────────────────┐
│  SPARQL Engine       │ ◄── QualiaDB :4242 | Wikidata | Both
└──────────┬───────────┘
           │ Returns semantic bindings
           ▼
┌──────────────────────┐
│  SmolLM2 / Llama 3   │ ──► Natural language answer + JSON-LD node
└──────────────────────┘

If the SPARQL query fails validation, the error is fed back to the model for self-correction (up to 3 retries). Token streaming uses inferWasmAsyncMeasured; CPU-WASM runs in a dedicated worker with packed SIMD128 Q8 kernels, while WebGPU yields around asynchronous device work.