Meet Metis AI — the brain inside every PressTwo agent
Frontier-class reasoning, trained on your business, wrapped in enterprise-grade security. Metis is the intelligence layer that turns a pile of your documents into a real 24/7 employee — one that can look up an order, book a slot, answer a policy question, and escalate cleanly when it's out of its depth.
Built on frontier AI infrastructure — including Anthropic's Claude — plus Voyage AI embeddings, pgvector retrieval, and Upstash Redis rate limiting.
What Makes Metis Different
Three things have to be true for enterprise AI to work
It has to be smart enough, honest about what it knows, and safe by design. Most AI products nail one of the three. Metis is built for all three.
Frontier Intelligence
Metis runs on a frontier-class large language model — the same class of model powering the most capable AI assistants on the market. It reasons, plans, asks clarifying questions, and cites its sources.
Your Knowledge, Not The Internet
Metis never guesses from public web data in front of your customers. It answers strictly from the documents, websites, databases, and APIs you connect — with citations back to the source.
Safe By Design
Constitutional AI training at the model layer. Prompt-injection defenses and tenant isolation at the application layer. SOC 2 Type II + ISO 27001 hosting at the infrastructure layer. Safety isn't a checkbox — it's baked into every surface.
How It Works
How Metis becomes an expert in your business
The loop runs every time a visitor sends a message. Five steps, all under five seconds.
Ingest
You upload files or paste URLs — PDFs, CSVs, Excel sheets, websites, FAQs, policy docs, API specs. Metis parses, cleans, and deduplicates everything.
Embed
Every passage gets converted into a 512-dimensional semantic vector using Voyage AI embeddings (voyage-3-lite). These live in Postgres with pgvector.
Retrieve
When a visitor asks a question, Metis vectorizes the query, finds the top 10 most similar passages across your knowledge base, and reads them — all in under 500ms.
Generate
The frontier LLM reads the retrieved passages plus your custom system prompt, your business-hours status, and any tool output. Then it writes a grounded, on-brand answer.
Verify
Every response is logged with the retrieved sources, token counts, and tool calls. You see exactly what Metis read before it replied — full audit trail.
This isn't fine-tuning — and that's deliberate.
Fine-tuning bakes your data into a model's weights, which means the next time you update a price or close a location, you'd have to re-train. Metis uses retrieval instead. Change a doc at 10:47am and it's in the next answer at 10:47am — no re-training, no waiting, no stale knowledge. And because your content never becomes model weights, it's never used to train the base model.
Capabilities
What Metis can actually do
Not a chat widget stapled to FAQ search. A reasoning engine that plans, looks things up, and knows when to punt to a human.
Retrieval-Augmented Generation
RAGEvery question triggers a semantic search over your knowledge. Metis reads the top-matching passages and writes an answer grounded in the retrieved text. No hallucination, no fabricated URLs, no made-up policies.
Tool Use & Real-Time Lookups
ToolsMetis can call your APIs, query Postgres or MySQL, read Google Sheets, pull live Shopify orders, book Calendly slots, and publish PushPress purchase links — all inside the conversation.
200k-Token Context Window
ContextEntire employee handbooks, state compliance manuals, multi-hundred-page product catalogs — Metis can hold all of it in a single conversation without losing the thread.
50+ Languages, Auto-Detected
i18nA Spanish-speaking customer gets a Spanish answer. A Japanese visitor gets a Japanese answer. No configuration — Metis detects the language and responds in-kind.
Human Handoff Awareness
HandoffMetis knows what it doesn't know. When a question exceeds its knowledge or a visitor explicitly asks for a human, it captures context and routes to your team via email, Slack, or webhook.
Business-Hours Aware
HoursMetis reads your configured schedule and timezone every turn. Off-hours, it won't promise a live callback in five minutes — it captures a message and sets the right expectation.
Security
Defense at every layer — model, app, infrastructure
Encryption everywhere
TLS 1.2+ in transit. AES-256 at rest. Database backups encrypted with keys we don't hold in plaintext.
Tenant isolation
Row-level security on every table. One customer's knowledge can never reach another customer's bot — enforced at the database, not just the app.
No training on your data
Your conversations, documents, and uploads are never used to train the underlying model. Metis reads them to answer; the model doesn't learn from them.
Prompt-injection defense
Untrusted content from tool results, scraped pages, and uploaded PDFs is treated as data — not instructions. Attacks that try to rewrite the bot's behavior through a malicious document don't work.
Rate limiting & abuse caps
Per-visitor, per-IP, and per-bot sliding-window limits prevent scraping, cost-amplification, and runaway usage. A hard safety cap at 5× plan pauses service before a misbehaving widget burns through your budget.
API-key authentication
External callers authenticate with a bot-scoped API key. Revoke one in a click; it can't be used to impersonate other bots on your account.
Compliance
Built for regulated industries
Healthcare, finance, legal, insurance — the categories where "we're kind of secure" doesn't fly. Metis inherits certifications from its infrastructure partners and adds the application-level controls you need to audit.
SOC 2 Type II and ISO 27001 certifications refer to the underlying AI infrastructure provider. PressTwo's own formal certifications are on the roadmap; request a security review from your account team for Enterprise deployments.
Reliability
The numbers behind the always-on promise
Observability
Nothing about Metis is a black box
Every conversation is captured, every tool call is logged, every lead capture is traceable. You can audit exactly what Metis retrieved, what it said, what your API returned, and where a conversation handed off to a human.
- Full conversation transcripts with timestamps
- Per-message token usage for cost attribution
- Tool-call input/output logs, scrubbed for PII
- Lead capture events with source attribution
- Webhook delivery receipts
- Handoff triggers and reasons
{
"id": "conv_9f3a...",
"bot": "Press Two AI",
"messages": [
{
"role": "user",
"content": "What are your hours?",
"at": "2026-04-19T14:22:11Z"
},
{
"role": "assistant",
"content": "We're open Mon–Fri 9–5 ET. Today we close at 5pm.",
"tokens_in": 1842,
"tokens_out": 23,
"sources": ["kb_policies.md:hours"],
"at": "2026-04-19T14:22:13Z"
}
],
"tool_calls": [],
"business_hours_status": "open"
}The Stack
What you're actually buying
Metis isn't one component — it's an opinionated stack tuned to do one thing well: answer your customers' questions in your voice, without hallucinating.
Deploy Metis on your site in under a minute
Start free, no credit card. Upload your docs, paste an embed snippet, and your first Metis-powered agent is answering real customers tonight.
Metis AI is a product of PressTwo, built on frontier AI infrastructure including Anthropic's Claude. Metis is the intelligence layer — the assembly of retrieval, tools, safety controls, and reasoning tuned for customer conversations. Anthropic and Claude are trademarks of Anthropic PBC.