Deep dives on RAG, security, and building AI infrastructure that respects your data.
Connectors, a self-organizing wiki, safe custom skills, self-masking documents, ceilings you control, and an assistant that reads the disagreement between documents: a one-page tour with every door linked to its guide. Request a demo at the end.
When your documents tell the same story differently, the assistant reads them TOGETHER - and tells you, in answer, where the versions disagree. A person decides what the honest version is. Request a demo.
The assistant reads the documents and answers where your people already work, and the systems you already run can both give tools to the assistant and reach its answers - without new code. Request a demo.
Usage ceilings in an AI workspace usually arrive as a report after the bill. Now they arrive before the spend: per person or for the whole install, checked before every answer. Request a demo and watch the ceilings hold.
Codici fiscali, account numbers, IBANs: documents entering the workspace mask those details by themselves, number by number - the staff keeps working while the machine never quotes the raw data. Request a demo.
Write the routine's instruction once; your people call it from the chat with a command like /translate. Nothing gets executed and nothing connects to the outside: the safe way to share your organization's way of working. Request a demo.
The knowledge your team archives now builds its pages, keeps the connections and flags where documents disagree - within rules your IT office writes. Request a demo and see the wiki assemble itself.
Your customers write on WhatsApp, Telegram, Discord and Slack: from today those same channels answer with what your documents say, citation included. Request a demo and watch your knowledge answer where your customers already are.
Per-area tour of the platform's gains since v0.22: connectors, wiki rails, custom skills, document DLP, hold-and-settle quotas, bidirectional MCP, synthesis. Each links its guide + post.
The synthesis pipeline: an archive's documents become combined wiki pages - budgeted runs, judged contradictions, human-approved writes only. Census: F-SYNTH-1.
Bidirectional MCP: the marketplace's external servers become agent skills while the built-in server exposes the workspace RAG to IDEs - fallback, health, pins. F-MCP-1.
How Simmetric Chat's quota engine covers tokens, storage, and real money - the ledger-backed no-second-meter design, the fail-open unset semantics, and the two-phase hold-and-settle reservation that turns spending limits from a report into a gate.
The document DLP: three detection tiers, document-wide numbered placeholders, a total re-embed, AES-256-GCM originals - and two explicit unmask gates. Census: F-DLP-1.
How Simmetric Chat's custom skills compile a stored prompt template into a spotlighted, untrusted context entry - the whitelist and single-pass rules, the three scoping levels, DLP masking before compile, and the max_skills freemium (3 or unlimited).
How Simmetric Chat's wiki engine governs archive writes with a three-layer instruction model - code-enforced hard rules, an advisory per-archive schema prompt, and structured fields - and what the precedence chain means in practice.
How Simmetric Chat routes external channels (Telegram, Discord, Slack, WhatsApp) through one adapter contract into the same grounded RAG pipeline - with rate and session limits protecting the workspace.
SSO with SAML 2.0 and OIDC, SCIM provisioning, an immutable audit log, white-labeling, encrypted backups, and air-gap licensing for regulated teams.
A practical guide to embedding Simmetric Chat on your site: sandboxed iframe loader, server-resolved model pin, lead capture, and CORS setup.
How Simmetric Chat authenticates every request, hashes API keys with HMAC-SHA256, signs webhooks and Web Push, and redacts PII with DLP filters.
Simmetric Chat scales horizontally behind any load balancer, with no sticky sessions. How state moves to Redis, Postgres, and the pg-boss job queue.
How vector similarity and PostgreSQL full-text search run in parallel and fuse through Reciprocal Rank Fusion into answers with citations and relevance scores.
A practical architecture for zero-outbound AI: Ollama as the local LLM, LanceDB for vectors, DLP redaction on every prompt, all in one offline container.
How Simmetric Chat's role-based access control uses 44 permissions and 15 menu sections to protect workspaces, projects, and audit trails.
The two-step loader and sandboxed iframe behind Simmetric Chat's widget, with isolated sessions, rate limiting, and layered knowledge-base access.
How Simmetric Chat orchestrates Ollama, OpenAI, Anthropic, OpenRouter, and Gemini with 20 provider presets, a Cmd+K model palette, and side-by-side comparison.
How Simmetric Chat reads scanned PDFs with a server-side vision-model OCR pipeline - page rendering, an Ollama vision model, and a human approve step.
The full deployment map for Simmetric Chat - a six-service Compose stack, a single all-in-one container for air-gapped hosts, and a one-click Coolify path.
Simmetric Chat speaks MCP in both directions - a curated marketplace of external tool servers, one-click installs, and per-chat control over what your agent can reach.
