Limits

Figures

• Upload and non-retaining: Stored upload sizeBytes maximum: 500 MiB (500 * 1024 * 1024 bytes); filename length: 1–500 characters; relativePath: up to 2,048 characters. Non-retaining file: 1,000,000 bytes; text: 250,000 characters; schema fields: 32; instructions: 8,000 characters; output: 32,000 bytes; deadline: 20 seconds.
• Connectors and Google Drive: HTTP source-ingest maxObjects: 1 to 50. Google Drive folders: up to 20 (subfolder depth up to 20); downloaded file: 25 MiB; exported Google Doc/Sheet/Slide: 10 MB; new files per import run: 500; listing continuation jobs: up to 50.
• Embeddings and multi-size chunking: Google gemini-embedding-001 input: 2,048 tokens (2,016 UTF-8 bytes) per text, batch size up to 100 at 3,072 dimensions. Multi-size chunking: up to 5 sizes total (mainSize plus up to 4 extraSizes), each 100–12,000 characters and at least 1.5x apart; default three_sizes preset is 500 / 1,500 / 4,500.
• Search, answers, and studied fields: Search read: 90 per 60 seconds; Search answer: 12 per 60 seconds. Studied field queries (field_answers_v1): up to 6 fields, 6 filters, 200 rows (default 50), and 50 groups; incomplete flagged below 90% field coverage. Table facts per document: up to 20,000 graph facts; retrieval tableEvidence: up to 24 facts and 12 findings.
• Lineage and Explore: Catalog map: up to 2,000 documents considered, 200 drawn, and 800 links. Investigation and Explore canvas: up to 200 nodes and 320 connections across 1–5 collections; focus: 40 items; undo: 30 steps.
• Context Engine and Context Studio: Per-document layer build: up to 500 documents per run (resume for the rest). Workflow Extract & Embed knowledge layers spend cap (capUsd): $0.01 to $10,000. Context sets: scope up to 8 collections or 20 documents, up to 100 approved entries, 20 rule templates, and 20 must-include documents/entities; up to 10 bound sets per agent credential. Structured SQL query templates: 1–1,000 rows, 1–30 s timeout, 1 MiB result cap, 4 concurrent runs per workspace.
• Ask history, Learning, and Agent memory: Ask conversations: up to 200 turns; standing notes: up to 20 per collection (600 chars each); retention: 1–365 days (default 90). Learning retention: 7–90 days (default 30). Agent memories: 1–1,000 chars text, up to 20 citations, 120 remembers per credential per hour, and 5,000 total memory versions per member and workspace.
• Agent Context API and MCP: Active credentials per workspace: 50 (1–365 days expiry). Context pack maxTokens: 256–32,000 over REST (8,000 on MCP); maxLatencyMs: 1,000–60,000 ms. MCP output bounds: 64 KiB structured JSON, 80 KiB text. Rate limits: 300 req/min per IP, 120 req/min (30 heavy) per credential, 600 req/min per workspace; concurrency: 4 per credential, 8 per workspace, 64 total.