How portfoliochat.ai actually works.

portfoliochat.ai sits as three layers above your portfolio management system: data ingestion, a portfolio intelligence layer that turns raw holdings into the structured context an AI can reason over, and an agent orchestration layer where specialised agents do the work. Each layer is separable, so source systems, context and agent logic can evolve independently.

Every request follows the same path. Scoping resolves tenant and portfolios, preprocessing loads the context downstream agents will need, and an orchestrator picks the route: a direct reply for simple turns, or a workflow run where a planner dispatches specialists and a synthesis agent writes the final answer. Each agent operates inside a clear boundary the runtime enforces, so the planner stays focused on planning and specialists stick to their job.

Two roles of agents run inside portfoliochat.ai. Orchestrator agents plan which specialised agents to call and see only the process information they need to steer, not the full answers. Specialised agents pick up the actual tasks, produce the output, and hand a compact process report back to the orchestrator.

Every agent is built from the same pieces: a base contract that defines what a turn looks like, versioned prompts pulled from an external store, a configurable model class, and a registry of tools the runtime exposes per agent. Prompts and model choices are configuration, so changes are reviewable and reversible without touching code.

Memory works on two timescales. A long-lived conversation summary keeps continuity across turns, and a short-lived per-turn workspace lets agents share intermediate results. Prompts stay compact, and context carries forward where it should.

Every turn is logged end-to-end in Langfuse: prompts, tool calls, data sources, model outputs and human overrides. A reviewer can replay any historical run with the exact inputs and final answer the user saw. The trace is the audit record, not just a debug aid.

Each agent and end-to-end flow has measurable benchmarks in Langfuse: factuality, restriction adherence, tone, latency, citations. The pipeline replays against fixed inputs, so prompt or model changes are scored before they reach production. This is also how model upgrades stay safe.

Model choice is configuration, not a code dependency. An agent declares the class of model it needs and a resolver picks the right provider for the deployment: frontier hosted models, smaller specialised models, or local models inside a customer's own infrastructure. No vendor lock-in, no rewrite when the landscape shifts.

portfoliochat.ai connects to portfolio management systems via API, ingests CSV and JSON, and pulls portfolio-aware market coverage from RavenPack. Internal research and CIO content can be added as documents. The integration layer normalises everything into the same data model.