Search — Universal Timeline Search + Freya · Kbatch · Hexcast

μ · mueee

Wall time

connectors 37 · results 0 offline | v2.7

Rail stream Local digest

Data staging lane: top-hit digest feeds AI Lens / weights; switch Model to Mμ for Mistral synthesis over the same grounded cards (completion = component handoff).

Stage: idle — run a search to stage results.

Grounded search + Mistral setup

Two lanes:

Connectors (below) — live grounded search; always on.
Rail stream (this panel) — Local digest summarizes staged hits from the last run.
Mμ synthesis — runs only when Model = Mμ and the Mμ proxy + agent fields are filled. Tuning: Mμ agent → Mistral tuning.
Liquid AI / LEAP / offline — connector Liquid AI / LEAP adds LEAP, model library, playground, and Hugging Face rows. Point the playground or offline gateway URL at your host (self-hosted LEAP stack or an OpenAI-compatible proxy); any API keys stay on that server, not in this page. Settings: Edge & Liquid AI (localStorage).
CORS — some APIs block the browser on localhost. Use a fetch proxy (e.g. set localStorage uvspeed-fetch-proxy-url) or open connector links in a new tab.

Search across Wikipedia, Grokipedia, arXiv, Sacred Texts,
PubChem, Wayback Machine, Open Library & more.

Works offline-first. No account needed.

Board

Weights

Heatmap

Export

Suggest

Typing

Context

Transcript

The keyboard lab and full spine table live on kbatch.html (tab Spine & I/O) — not embedded here.

Video hex broadcast + spine table: hexcast.html (section Spine & I/O map).

Hub — where each surface gets data in/out and what usually blocks speed. Grouped; *.html in web/ (~115 files). Spool spec: plans/gridlock-throughline-spine.plan.md.

Group	Pages	Data in / out	Speed / blocker

Rows generated once at load. “Broken” = missing unified export packet, no BC wire, or CORS/network — not always a bug. T5 classify ~3ns/line local; connector fetch ~200–800ms RTT.

Strata & Buckets

Mμ agent (Mistral) — instructions & guardrails

Synced with mueee footer Model · fields persist in this browser for API / bridge calls.

Required for live Mμ replies: a proxy URL (mu-mistral-proxy-url or localStorage mueee-mistral-chat-url) and a non-empty instructions block (override or default from mueee-mu-context.js). Without both, the bridge shows an error string in Mμ notes.

Instructions override (optional; default = MueeeMu from mueee-mu-context.js)

Guardrails

Tone

Knowledge / context notes

Mistral proxy URL (Worker or your backend; no API key in browser)

Uses main prompt box above · deploy cloudflare/mueee-mistral-proxy-worker.js

Mistral tuning — sampling, lane, structured output

Merged into POST …/chat/completions by the bridge. Your Worker forwards the JSON body as-is. Tune creativity vs focus here; guardrails in the block above define your policy lane in prose — API guardrails arrays are optional below for server-side category blocks.

Temperature

Top P (optional)

Max tokens (optional)

Safe prompt (moderation) JSON object output

API guardrails JSON (optional; advanced)

Chat completions
Usage (prefix, safe_prompt, guardrails field)
Prompting
Sampling (temperature, N, top_p, penalties)
Structured outputs · Custom · JSON mode
Function calling · Agents + tools
Vision
Custom guardrails (lane / moderation)

Mistral Document AI — OCR · QnA · annotations

Same Worker can expose POST /api/mistral-chat and POST /api/mistral-ocr → https://api.mistral.ai/v1/ocr. Document QnA uses chat completions with a document_url part (no key in the browser).

Document AI overview
Document QnA
Annotations (structured extraction)
OCR processor
Cookbooks: batch OCR, tool usage, data extraction, document understanding

OCR proxy URL (optional; defaults from chat URL → …/api/mistral-ocr)

Public document URL (PDF) for tests — must be reachable by Mistral’s API

QnA uses the main prompt above as the question

Transcript → wire (multi-pass prompts)

After OCR or ASR, run phases in order in Mμ, Gemini, or Mistral. Replace placeholders, then copy. Ingestion tiers: docs/search-ingestion-wires.md.

Phase 1 — Janitor

Phase 2 — Wire brief

Phase 3 — Fact-check

Verbatim + diarization (short video)

Fetch proxy (CORS) — full connector power

Deploy cloudflare/uvspeed-fetch-proxy-worker.js. Paste the Worker URL (must end with /api/fetch-proxy). Fixes arXiv, Invidious/YouTube, and strict-browser fetch — same engine as Google alternatives, without their tracking.

Proxy base URL

Edge & Liquid AI (LEAP, playground, offline gateway)

This shell does not store vendor secrets in localStorage (same idea as the Mistral Worker). You can run Liquid / LFM / LEAP self-hosted or behind your reverse proxy: put API keys and auth only on that host, then set the playground base URL or offline gateway below to your origin (e.g. https://playground.your.internal/chat or http://127.0.0.1:8080). The Liquid AI / LEAP connector still adds public link-outs to LEAP, the model library, optional Liquid cloud playground, and Hugging Face.

Liquid playground base URL

Playground model id

Your offline / local LLM base URL (optional)

Ranking & source governance (fresh · reliable · longevity · bias)

Total score blends freshness (dated items decay), reliability (factuality + transparency), and longevity (institutional). Political bias labels are curated heuristics (Ground News–style UX) — verify independently; expand registry in history-search-engine.js or your Worker.

Bias preference

Document

Content

Analysis

References

Patterns

Click a search result to view document analysis.