Back office ops · Production

Dropbox Dash: Building RAG and multi-step AI agents for enterprise knowledge management

The problem

Knowledge workers face information scattered across multiple applications and formats, making document retrieval tedious and time-consuming; data fragmentation hinders collaboration and productivity and creates costly security risks.

First attempt

RAG alone is incapable of performing complex, multi-step business tasks that require domain knowledge, contextual information, and multi-stage planning and execution.

Workflow diagram · grounded in source
1
User query submitted
trigger
“Show me the notes for tomorrow's all-hands meeting”
2
RAG content retrieval
ai_action
“we landed on a traditional information retrieval (IR) approach combined with on-the-fly chunking and reranking”
3
AI agent planning via DSL
ai_action
“The planning stage involves breaking down a user's query into a sequence of high-level steps. This is done by an LLM, which interprets the query and generates simple code statements to express the logic of responding to the user's query.…”
4
Code validation and repair
validation
“The code is validated through static analysis to ensure correctness, safety, and to detect missing functionality. We intentionally allow the LLM to assume that missing functionality exists. If missing functionality is identified, we use …”
5
Step-by-step code execution
ai_action
“Having run-time type enforcement helps ensure that the data and objects being operated on are the types of values that we expect.”
6
Final response returned
output
“The final response to the user's query is the result of the last step. In this example, the list of documents will be returned to the user.”
Reported outcome

The integration of RAG and AI agents significantly enhanced Dropbox Dash, achieving high-quality results in under 2 seconds for over 95% of queries.

Reported metrics
Query response latencyunder 2 seconds for over 95% of queries
Reported stack
Dropbox DashRAGLLMDSL
Source
https://dropbox.tech/machine-learning/building-dash-rag-multi-step-ai-agents-business-users
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The integration of RAG and AI agents significantly enhanced Dropbox Dash, achieving high-quality results in under 2 seconds for over 95% of queries.

What tools did this team use?

Dropbox Dash, RAG, LLM, DSL.

What results were reported?

Query response latency: under 2 seconds for over 95% of queries (source-reported, not independently verified).

What failed first in this deployment?

RAG alone is incapable of performing complex, multi-step business tasks that require domain knowledge, contextual information, and multi-stage planning and execution.

How is this back office ops AI workflow structured?

User query submitted → RAG content retrieval → AI agent planning via DSL → Code validation and repair → Step-by-step code execution → Final response returned.

Related back office ops cases