Skip to main content

Multi-Turn Followup Capabilities

This document outlines the system's capabilities for handling multi-turn conversations, ensuring context is preserved, switched, or recovered intelligently across user interactions.

Overview​

The system supports 8 core scenarios designed to handle natural conversation flows, from simple "show more" requests to complex context switching and reference resolution.

Visualizing the Flow​

Context Preservation & Refinement​

This flowchart illustrates how the system maintains context when a user asks for more results or refines their search (e.g., adding a budget).

Context Switching (Anti-Poisoning)​

The system detects when a user wants to change topics completely, preventing old context (like "books") from polluting a new search (like "toys").

Core Capabilities​

1. Context Preservation ("Show More")​

Goal: Keep the current search active while fetching more results or applying minor filters.

CapabilityExampleBehavior
Preserve Type"Show more"Keeps showing the same product type (e.g., grills).
Preserve Category"More please"Maintains the current category context.
Add Budget"Under 30 euros"Applies budget limit to the existing search.

2. Progressive Context Building​

Goal: Allow users to build a complex query step-by-step.

Example Flow:

  1. "Gift ideas" (Generic)
  2. "For mom" (+Recipient)
  3. "She likes gardening" (+Hobby)
  4. "Under 50 euros" (+Budget)

Result: A search for "Gardening gifts for mom under 50€".

3. Anti-Poisoning (Clean Pivots)​

Goal: Detect when the user changes their mind and wants to start over.

  • Triggers: "Instead...", "Actually...", "I want X instead".
  • Action: The system clears product_type, category, and other specific constraints to avoid "leaking" old context into the new search.

4. Agile Context Switching​

Goal: Support rapid switching between topics and returning to previous ones.

  • Switch: "Actually, show me books." (Switches to Books)
  • Switch Back: "Back to the games." (Restores Games context)

5. Reference Resolution​

Goal: Understand pronouns and references to authors or products.

Supported Patterns:

  • Pronouns: "his works", "from this author", "it", "that one".
  • Ordinals: "the first one", "the second product".
  • Implicit: "How much is it?" (Refers to the last shown product).
  • Mixed Intent: "Is this suitable for an 8-year-old?" (Pronoun + age context).
Deterministic Detection

Ordinal and pronoun detection uses regex-based patterns for Estonian language, running at PRIORITY -2 (earliest) in the routing pipeline. This ensures ~2ms detection instead of LLM latency. See Product Inquiry Detection for implementation details.

6. Budget Management​

Goal: Flexibly handle budget constraints.

  • Add: "Under 20 euros"
  • Modify: "Actually up to 50"
  • Remove: "Price doesn't matter"

7. Constraint Overrides​

Goal: Handle negative constraints (exclusions).

  • Exclude: "Not books"
  • Re-include: "Actually books are okay"

8. Mixed Intent Recovery​

Goal: Handle questions in the middle of a search flow without losing the search context.

  1. Search: "Show me games."
  2. Question: "Is this one suitable for 5 year olds?" (Context preserved)
  3. Resume: "Show more." (Resumes game search)

Technical Implementation​

The context is managed via lastSearchParams, which persists critical fields across turns:

  • queryForSearch: The base query.
  • product_type / csv_category: Core classification.
  • budget: { min, max }.
  • recipient / ageGroup: Gift context.
  • isContextSwitch: Flag to indicate a hard reset.
  • Test Suite: qa-surface/test-multi-turn-followups.ts
  • Types: app/api/chat/types/index.ts (ChatRequest.lastSearchParams)
  • Show More Utils: app/api/chat/utils/show-more.ts (LastSearchParams)
  • Parallel Orchestrator: app/api/chat/orchestrators/parallel-orchestrator.ts (metadata construction)
  • Context Types: app/api/chat/context/types.ts (GiftContext)
  • Product Inquiry Detection: app/api/chat/orchestrators/context-orchestrator/llm-query-router.ts