Deploy and operate an embedded workflow

Deploy Foliqant as part of the Python application that calls it. It is an in-memory library, not a standalone queue or HTTP service. The host controls incoming transports, authentication, authorization, persistence, idempotency, and process supervision.

Package the configuration and dependencies

Use Python 3.12 and install the reviewed wheel plus only the adapter extras your configuration needs (openai, anthropic, azure, google, bedrock, mcp, and/or telemetry). The repository's tested set is locked in uv.lock; see installation. Ship config/settings.yaml, workflow definitions, local schemas, and prompt files together as one reviewed artifact. Paths resolve beneath the configuration root, so include their relative layout. Do not put credentials in that artifact.

Set required environment variables in the host environment or adjacent config/.env. Explicit $NAME references in deployment fields resolve when the application opens; process values take precedence over .env. Preparation can validate structure offline, while missing required environment values fail at startup. Configure environment explains the supported fields and resolution rules.

Open once and shut down cleanly

At startup, call prepare_application(config_path, handlers=...), then enter open_application(prepared, environment=os.environ, plugins=...) once per process. Register trusted handler implementations and tool authorizers in application code; configuration cannot import Python functions. Hold the async context while requests are accepted and leave it during service shutdown. The context drains owned work before closing clients. A forced process kill loses unfinished in-memory executions, so the host needs durable coordination if recovery after a crash matters.

Configured admission capacity, deadlines, step visits, model/tool attempts, and collection item caps bound work within one process. They are not global rate limits or service-level guarantees. Scale-out hosts need their own shared admission policy if that matters. See Configure limits.

Observe safely

The result carries a run ID, configuration revision, per-flow records, elapsed times, and measured usage. Token measurements can be null when a provider did not report them. Do not add parent and child usage values together. Optional telemetry emits restricted labels; keep payloads, prompts, identities, credentials, raw exceptions, and customer evaluation data out of logs. Store full results or evaluation reports only under your application's data policy. See observability for telemetry setup, runtime configuration for exact fields, and Handle errors for reconciliation and retry decisions.