SCRAPPY_RAG
local · $0
local question desk

Ask what the
corpus knows.

OpenRouter model and provider facts, grounded in the local Scrappy store. The reader cites what it used and says when the answer is not here.

1024-dimensional Qwen3 embeddings
Local Qwen2.5-3B reader
No paid API calls

Question

POST /api/chat
Retrieval mode
Loading the model on first use.
The 2 GB local reader wakes lazily; later answers are faster.
Warp replayRe-running a past Ask with its stored settings.

Evidence strip

retrieval preview · not the answer

Grounded answer

Retrieved context

Question log

tap a question to expand · one-click Replay · copy
corpus ledger

The store,
in the open.

Browse every chunk available to retrieval. Counts come from the live LanceDB store, not from this page.

Canonical catalog facts
Provider documents
Rendered page excerpts

Chunks

loading

Evidence trail

tap a row to read
retrieval anatomy

The corpus,
from above.

A PCA projection of 1024-dimensional embeddings. Nearby dots speak similar language; the map is an orientation aid, not a distance claim.

Hover or tap a point
Target a question
Inspect its projected neighborhood

modelproviderpagequeryexplored
— points
move to inspect · scroll to zoom · drag to pan · fog lifts on explored chunks
projecting store
Fog unavailable — showing fully-lit map (feedback history could not be loaded).

Target me

The white cross is your question. Amber rings mark its nearest neighbors in this 2D projection.

point inspector

Move across the map

Titles and document kinds appear here.

mirror database

The data,
laid bare.

A read-only dashboard over the Scrappy mirror SQLite store. Charts and tables reflect live API responses — nothing is cached on this page.

Catalog overview
Time-series activity
Interactive table explorer

Tables

Select a table

Choose a table from the left rail to explore its rows.
feedback desk

Review and rate,
improve the corpus.

Rate answers after they run. Review past interactions with full detail and history. Filters and pagination supported.

Newest first
Full context text
Model fingerprints

Review desk

tap to expand
model registry

The registry,
on the record.

Capability facts straight from the model registry API: context windows, pricing, modalities, reasoning support, rate limits, and benchmarks. Numbers come verbatim from the registry — nothing is invented here.

Lookup by name
Filter by capability
Everything as-of the registry

Model lookup

GET /api/registry/search
type to search · top 8tap a row to open its card

Filter search

GET /api/registry/search
Input modalities
tap a row to load it above
mirror health & pulse

The mirror,
always fresh.

Hourly public surface mirror of OpenRouter catalog, pricing, rankings, and benchmarks. Read-only view into the local harvest engine, request budgets, and fetch ledger.

Self-hosted Python engine
Append-only fetch ledger
Zero paid API calls

Connecting to mirror database and health endpoints…

Engine Vital Signs

Cycle & Freshness

Cycle State

Last Cycle ID
Last Run Time
Finished Time
Killswitch

Freshness & Lag

Mirror Event
Store Chunks
Freshness Lag
Health Note

Combined Budget

Nominal
Hour Usage
Day Usage
Total Ledger
Ledger Failures

Budget Limits & Surface Counters

S1–S6 Surface Budgets
Surface Class Description Hour Use Hour Cap Day Use Day Cap Usage Bar
Loading surface budgets…

Store Inventory Counts

Local Mirror Tables
Catalog Models
Price Changes
Benchmark Scores
Total Events

Fetch Ledger Tail

Last 20 HTTP Requests
ID Timestamp Class Surface Status Bytes Latency URL
Loading ledger tail…
Full database inspection, SQLite schemas and per-day distributions: Open Data Explorer →
live deployments

Scrappy-powered
websites.

Sites running in the wild on Scrappy data. Each card links straight to the live deployment and names the host that serves it.

Two live deployments
Built on the same store
Links open in a new tab

Directory