Built for depth.

Whether it's a single idea or a research lab with millions of pages, everything stays fast. No loading screens. No waiting. Just your work, flowing.

An editor with no bottom.

Most editors slow down after a few hundred pages. They stutter, freeze, and eventually crash. Draftbox doesn't. Scroll from page 1 to page 100,000 and it feels exactly the same — no lag, no freezing, no ceiling. Write a paragraph or write an encyclopedia. The editor treats them the same.

Pages	Others	Draftbox
500	Smooth	Smooth
1,000	Laggy	Smooth
Infinite	Impossible	Smooth

Reads like a library. Writes like a scribe.

The AI can write hundreds of pages directly in the editor while you watch — word by word, grounded in your data, with zero deviation from your sources. And the editor keeps up without breaking a sweat.

Live writing

AI writes live in the editor — hundreds of pages, word by word, right before your eyes
Every claim traces back to your sources — zero deviation from your data
Words stream smoothly at the frame level — no stuttering, no flicker, no jitter
Paste 50,000+ characters at once? It inserts progressively in batches of 5 — no freeze
Over 50 cards on screen? Virtualization kicks in automatically — stays smooth
Summaries generated on-the-fly for 32 blocks at once, kept for 30 minutes *

Processing & indexing

When you change a block, only that block gets re-processed — everything else is remembered
Upload the same file twice? It skips ~30% of the content automatically *
Blocks you haven't changed cost zero to re-index *
32 blocks processed at the same time on Scientist plans *
IconDatabase writes happen 50 blocks at a time — no round-trips per block
Grounding score computed in real-time — the AI knows when to stop and when to dig deeper

Connected users. Multiplied performance.

When multiple users collaborate in the same workspace, the engine searches everyone's content in a single pass — not one search per person. Duplicates across users are eliminated automatically before they ever hit the database. More people, more knowledge, same search time.

One search across all connected users — no matter how many people are in the workspace, it's always one query, not one per person
Automatic deduplication: when two users upload the same file, only one copy gets indexed — the other is skipped instantly *
64KB Bloom filter per workspace catches duplicates before they ever hit the database — 7 hash functions, <1% false positives *
32 blocks processed in parallel on Scientist plans (64 on Scientist Lab) — everyone in the workspace benefits from the same throughput *
More users means more knowledge, but search time stays nearly flat — the engine works smarter, not harder
Summaries generated on-the-fly for 32 blocks at once and kept for 30 minutes — shared across everyone in the workspace *

Read query routes. Write free.

Every query creates a ephemeral read-only query route of your workspace. Writes never wait for reads. Reads never block writes. Under load, the engine spawns extra query routes to absorb the burst — and destroys them when the wave passes.

Read / write isolation

Queries hit read-only query routes — writes hit the primary, unblocked *
Dedicated connection pools per query route — no pool contention *
Burst absorption: extra query routes spawn automatically under load, round-robin routed *
Pre-warmed query routes ready before predicted traffic spikes *
Query-route-aware caching — each query route gets isolated cache keys, no cross-contamination *

Reliability & compliance

Zero-downtime schema migrations — query route serves old schema while primary upgrades *
Disaster recovery — if the primary fails, a query route is promoted to temporary primary *
GDPR cascade — right-to-be-forgotten destroys all query routes containing user data *
Per-query-route encryption keys — if one key is compromised, others stay safe *
Forensic audit replay — reconstruct exactly what a user saw via a query route *
Governor self-healing — stuck query routes are detected and destroyed automatically *

Multi-embedding precision pyramid

Query routes don't just isolate reads — they also unlock tiered embedding precision. Instead of one approximate search, the engine runs a 3-stage pyramid: coarse binary filter, summary re-rank, then full-precision scoring.

Binary

Hamming distance — fastest, coarse filter

Pro+

Summary

Halfvec — re-rank with summary embeddings

Business+

Essence

Halfvec — distilled semantic essence

Enterprise+

Full

Full halfvec — maximum precision

Scientist+

Words on screen before you blink.

The time between pressing enter and seeing the first word matters. Draftbox cuts that gap to almost nothing — and keeps the stream smooth from start to finish.

Speculative start

The AI starts answering before the full pipeline even finishes. Budget, routing, and cache checks happen in parallel — so you see words faster.

Smooth streaming

Words arrive smoothly, not in stuttering bursts. The stream is chunked at the word level and rendered frame-by-frame — no jitter, no flicker.

Hybrid search that finds what you mean.

Two search engines run in parallel — BM25 for exact keyword matches and vector search for semantic meaning. Results are merged with Reciprocal Rank Fusion, reranked with cross-attention, and filtered by source, date, status, and type. Your knowledge, found — not just matched.

BM25 + Vector + RRF

BM25 full-text search with tsvector — exact keyword matching with term frequency scoring
Vector semantic search with halfvec(1024) — finds meaning, not just words
Reciprocal Rank Fusion (k=60) merges both rankings — best of both worlds
Cross-attention reranker (Cohere) re-scores top results for precision
LLM-based query expansion — your search terms are enriched before retrieval

HNSW + Faceted Filters

HNSW indexes (m=16, ef=64) for sub-linear search time at scale
Faceted filters — source provider, document status, tool ID, date ranges
Binary prefilter — Hamming distance coarse filter before full-precision scoring
Click-boost signals — items you open more rank higher over time
Time-travel queries — search what you knew at a specific moment

Retrieval pipeline

Query input→Query expansion (LLM)→BM25 search→Vector search (HNSW)→RRF fusion (k=60)→Faceted filters→Cross-attention rerank→Click-boost→Results

It doesn't crash. It adapts.

Under extreme load, most systems fail. Draftbox degrades. The governor monitors every service in real-time and adjusts the system across three tiers — so your work keeps flowing, even when the infrastructure is under siege.

Normal

Full LLM capacity. All features active. Every service running. Circuit breakers closed. This is where the system lives.

Stressed

Reduced LLM params. Non-critical features paused. Circuit breakers monitoring closely. Core functionality stays — extras wait.

Emergency

Minimal LLM. Only critical features. Circuit breakers open. The system survives — and recovers automatically when the storm passes.

Circuit breakers

Per-service failure tracking with configurable thresholds
Open state kills non-critical services automatically
Half-open state tests recovery before reviving
Closed state restores full operation after success

Self-healing

Stuck database query routes detected and destroyed automatically
Failed integrations retried with exponential backoff
Health scores per integration — visible in the UI
Auto-recovery when conditions return to normal

Speed is not a luxury. It is respect for your time.

Enter Performance by plan Compare plans

See what each plan unlocks.

Query route counts, embedding precision, encryption, disaster recovery — all gated by plan.

Compare plans