> ## Documentation Index > Fetch the complete documentation index at: https://docs.reflections.ai/llms.txt > Use this file to discover all available pages before exploring further. # Architecture decision records > Index of all architecture decisions in the Reflections platform. Architecture decision records (ADRs) capture the reasoning behind significant technical choices in the Reflections platform. Each record describes the context, the decision, alternatives that were considered, and the consequences of the chosen approach. ADRs are numbered sequentially. Superseded decisions remain in the index for historical context, with their status clearly marked. ## Core architecture Foundational decisions about repository structure, runtime standards, and the two-plane execution model. Enforce clear dependency direction and reduce cross-package coupling in a multi-app system. Standardize runtime and build behavior to minimize environment drift and packaging surprises. Separate low-latency inference from asynchronous learning to preserve reliability and control risk. ## Data and storage How the platform stores, evolves, and retrieves data including temporal facts, query segregation, and vector retrieval. Use Supabase Postgres with pgvector and pg\_trgm for transactional, relational, and vector workloads. Preserve auditable memory evolution through temporal facts and gated patch promotion. Constrain data mutation capabilities by design-time import boundaries. Build evidence retrieval that combines semantic recall with graph context while remaining low-latency. ## Security and auth Authentication, authorization, tenant isolation, and content safety across the platform. Support both service-to-service and end-user auth while enforcing server-derived reflection roles. Keep API boundaries explicit with consistent route-level authn/authz and error semantics. Define defense-in-depth tenant isolation across API authorization and database row-level policies. Allow public reflection access via constrained anonymous sessions without exposing privileged surfaces. Protect ingestion from hostile or oversized source inputs while preserving deterministic pipeline behavior. Canonicalize team-sharing onto Clerk Organizations using free-plan-compatible roles while preserving creator-private access semantics. ## Operations CI/CD pipelines, observability, migration governance, and ingestion orchestration. Make source ingest retry-safe and recoverable while preserving evaluation-gated learning. Provide actionable diagnostics without making core flows brittle. Keep release quality high through automated, ordered validation gates. Keep database schema evolution deterministic and verifiable across environments. ## Frontend and mobile Client-side architecture, voice runtime, and mobile API contracts. Balance server-state correctness, local UI responsiveness, and realtime ingest visibility. Replace self-hosted voice agent with a managed conversation provider while preserving knowledge-graph authority. Expo mobile configuration contracts in a pnpm monorepo. Superseded by native Swift iOS app. Centralize mobile-facing DTO schemas via shared Zod schemas. Superseded by native Swift iOS app. Explicit three-mode session routing (interview, reflection, share) with mode-aware tool permissions and dynamic variable budgeting. Persisted prompt-pack authority, audience-scoped public runtime, gated artifact promotion, and freshness-aware public agent selection. ## Quality and governance Lint enforcement strategy, rate limiting, content monitoring, and domain taxonomy. Mechanically enforce architectural invariants through layered, defense-in-depth tooling. Keep the in-process limiter for now with explicit key derivation and clear migration triggers. Ingest daily Loti content-monitoring results into Postgres for talent protection workflows. Provide user-facing domain structure over the knowledge graph with hybrid classification and dual scoring. Truth-first relational atlas with disclosure-first shell, bundled relationship edges, and deterministic structure-preserving sampling. ## Voice cloning Voice clone pipeline, lifecycle authority, and readiness gating. Voice cloning via outbox pattern with quality gating and metrics tracking. Superseded by ADR-0029. Durable attempt-driven lifecycle with sample-first product contract and multi-condition readiness gate. ## AI strategy Vendor selection, model strategy, and AI capability boundaries. Separate inference, embedding, and voice-runtime vendor responsibilities with configurable provider selection.