Workflow · Production

Aimpoint Digital builds AI agent travel itinerary system on Databricks using multi-RAG architecture

The problem

Planning a trip is time-consuming and overwhelming: travelers spend over 5 hours researching and visit up to ~270 web pages before finalizing activities. Standalone LLMs like ChatGPT suffer from a recency issue — they may recommend closed or outdated venues — and are prone to hallucination.

First attempt

Standalone GenAI tools like ChatGPT generate itineraries that can be misleading or incorrect because the underlying LLMs lack up-to-date information, suffer from a recency issue, and may hallucinate inaccurate details such as recommending a restaurant that has since closed.

Workflow diagram · grounded in source
1
Traveler submits parameters
trigger
“Some of these parameters are destination city, destination country, dates of travel, travel purpose (business, leisure, recreation, etc.), travel companion(s) (friends, partner, solo, etc.) and budget.”
2
Query vectorized
ai_action
“The user query is converted to a vector using the embedding model”
3
Parallel RAG retrieval
ai_action
“Three simultaneous calls are made to the different RAGs in parallel to gather recommendations. On average, our system is configured to retrieve three places or events and three restaurants daily (breakfast, lunch, and dinner).”
4
LLM synthesizes itinerary
ai_action
“we use an LLM to combine the recommendations into a cohesive itinerary. We are using open source LLMs like DBRX Instruct and Meta-Llama-3.1-405b-Instruct on Databricks using Provisioned Throughput Endpoints with built-in guardrails to pr…”
5
LLM-as-judge quality check
validation
“We used an LLM to evaluate travel itineraries for professionalism. This is an automated way to evaluate responses from AI agent solutions without ground truths.”
6
DSPy prompt optimization
feedback_loop
“DSPy uses an LLM-as-a-judge in conjunction with a custom-defined metric to evaluate responses based on a ground truth dataset”
Reported outcome

The tool demonstrated transformative potential in the travel industry and received overwhelmingly positive feedback from stakeholders who appreciated the seamless planning experience and accuracy of recommendations.
Itineraries are generated within seconds.

Reported metrics
Traveler trip research timeover 5 hours
Web pages visited before finalizing trip~270
Itinerary generation timewithin seconds
Stakeholder feedbackoverwhelmingly positive feedback
Reported stack
Databricks Data Intelligence PlatformDatabricks Mosaic AI Vector SearchDBRX InstructMeta-Llama-3.1-405b-InstructDSPyDelta tablesMLFlow
Source
https://www.databricks.com/blog/aimpoint-digital-ai-agent-systems
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The tool demonstrated transformative potential in the travel industry and received overwhelmingly positive feedback from stakeholders who appreciated the seamless planning experience and accuracy of recommendations.

What tools did this team use?

Databricks Data Intelligence Platform, Databricks Mosaic AI Vector Search, DBRX Instruct, Meta-Llama-3.1-405b-Instruct, DSPy, Delta tables, MLFlow.

What results were reported?

Traveler trip research time: over 5 hours; Web pages visited before finalizing trip: ~270; Itinerary generation time: within seconds; Stakeholder feedback: overwhelmingly positive feedback (source-reported, not independently verified).

What failed first in this deployment?

Standalone GenAI tools like ChatGPT generate itineraries that can be misleading or incorrect because the underlying LLMs lack up-to-date information, suffer from a recency issue, and may hallucinate inaccurate details…

How is this workflow AI workflow structured?

Traveler submits parameters → Query vectorized → Parallel RAG retrieval → LLM synthesizes itinerary → LLM-as-judge quality check → DSPy prompt optimization.