Skip to content
Agile Defense

Curriculum / Foundry Foundations

ADVANCEDUnit 8day-08

AIP Chatbot Studio Assistant

not started~360 min
Configure an AIP Chatbot (formerly AIP Agent) with governed context and safe tools to support analysts.

Introduction

Scenario: New analysts need a guided assistant that explains risk context, asks clarifying questions, and recommends next steps.

Day 8 puts the analyst in the seat of a builder: you will assemble a guided decision assistant in AIP Chatbot Studio, the environment for building interactive assistants known as AIP Chatbots that are equipped with enterprise-specific information and tools. Throughout the disruption-response problem, new analysts have struggled to interpret supplier risk context, know which clarifying questions to ask, and decide what to do next. The chatbot you build today closes that gap by turning the Ontology, retrieval context, and a small set of governed tools into a conversational partner that explains risk, asks for missing detail, and recommends next steps.

This matters because a disruption response only works if frontline analysts move quickly and correctly. AIP Chatbots are powered by large language models, the Ontology, documents, and custom tools, supporting dynamic, context-aware read and write workflows. The behavior of the assistant is shaped by its instructions, which (together with tool descriptions and variable descriptions) are compiled into the raw system prompt for the LLM, while retrieval context fetches relevant content per message and application state variables customize and control how the model behaves.

Critically, this is governed AI, not a generic chatbot. AIP Chatbots are built on the same rigorous security model that governs the rest of the Palantir platform, and these platform security controls grant an LLM access only to what is necessary to complete a task. That means the risk context your assistant surfaces and the actions it can take are bounded by the permissions of the user who is chatting, and every tool you add inherits that same governance.

Capability focus: AIP Chatbot Studio — chatbots, retrieval context, tools, Workshop integration · Artifact: Supply-chain decision assistant

Key concepts

  • AIP Chatbot Studio and AIP Chatbots: the studio is where you build interactive assistants (AIP Chatbots) equipped with enterprise-specific information and tools, deployable inside Foundry, in Workshop applications, and externally via the Ontology SDK and platform APIs.
  • Instructions and the system prompt: a chatbot's instructions, tool descriptions, and variable descriptions are compiled into the raw system prompt for the LLM; effective instructions lead with an overview of the task, then the data and guidance on using application state and tools.
  • Retrieval context: information retrieved in response to each message and included with the LLM request, fetching relevant content from configured data sources (such as the Ontology and documents) so answers are grounded rather than invented.
  • Application state / variables: application variables referenced within prompts that customize and control the LLM's behavior, letting the assistant adapt dynamically to the current task or selected supplier.
  • Tools (six documented types): Action, Object Query, Function, Update Application Variable, Command, and Request Clarification — the LLM-callable capabilities that let a chatbot read Ontology objects, run Foundry functions or AIP Logic, edit data, or pause to ask the user for more information.
  • The platform security model: AIP Chatbots run on the same rigorous security model as the rest of Foundry, granting the LLM access only to what is necessary, so every tool and context source is permission-aware by default.

Companion video

Speedrun: Your First Agentic Workflow · open on YouTube

Hands-on activity

each step validates · the unit completes when all steps pass
  1. 1

    Define the chatbot role and boundaries

    Begin in AIP Chatbot Studio by writing the chatbot's instructions, which are compiled into the raw system prompt for the LLM alongside your tool and variable descriptions. Lead with an overview of the task — here, helping a new analyst understand supplier and shipment risk during a disruption — then state what the assistant should and should not do, such as explaining risk drivers and recommending next steps rather than committing to decisions. Setting these boundaries up front keeps the assistant focused on guidance and ensures its persona stays consistent across every conversation.

    not startedself-attested

    Self-attested: instructions are specific and escalation rules are set.

  2. 2

    Configure governed, permission-aware context

    Add retrieval context so the assistant fetches relevant content from configured sources — your supplier, shipment, and risk-assessment object types in the Ontology, plus any disruption playbook documents — and includes it with each LLM request instead of guessing. Because AIP Chatbots are built on the same rigorous security model as the rest of the platform, this context is permission-aware: the LLM is granted access only to what is necessary, scoped to the chatting analyst's own permissions. You can also use application state variables (for example, the currently selected supplier) to dynamically control and customize the assistant's behavior.

    not startedself-attested

    Self-attested: context is relevant and permission-aware.

  3. 3

    Add safe tools (object query / function)

    Equip the assistant with a small, governed toolset from the six documented tool types: an Object Query tool to let the LLM read specified object types with properties configured for filtering, aggregation, inspection, and traversal of links so it can pull a supplier's live risk score and linked shipments, and a Function tool to call a Foundry function or published AIP Logic workflow that computes a recommended next step. Keep tools read-oriented and minimal so the assistant explains and recommends without making unintended edits. Each tool inherits the platform security model, so it can only return or act on data the user is already permitted to see.

    not startedself-attested

    Self-attested: the chatbot has appropriate, constrained tools.

  4. 4

    Confirm the decision chatbot exists

    Validate the assistant end to end by chatting as a new analyst would: ask about a specific at-risk supplier and confirm the chatbot grounds its answer in retrieved Ontology context, uses the Object Query and Function tools, and recommends a clear next step. Test the Request Clarification tool by sending an underspecified prompt and verifying the assistant pauses to ask for the missing detail rather than guessing. Confirm that responses respect the chatting user's permissions, then save the chatbot so it is ready to be surfaced to analysts, including embedding it in a Workshop application.

    not startedinstance check

    Confirms the chatbot via the preview AIP Agents API; blocked (not failed) if the preview API is unavailable.