DeepRAG: Advancing enterprise AI agents from sparse retrieval to agentic, multi-document synthesis

For years, the promise of AI in the enterprise has been tantalizingly close. We've had systems capable of information retrieval, but retrieval does not equate to comprehension. Standard retrieval-augmented generation (RAG) architectures rely on simple vector or sparse retrieval to identify relevant document chunks for single-shot generation. This is inherently limited in multi-source retrieval and synthesis (MSRS) tasks where cross-document reasoning and stateful knowledge consolidation across massive, siloed document sets is required.

Today, we're closing that gap. We are thrilled to introduce DeepRAG, UiPath's advanced AI system that moves your agents from rudimentary fact-finding to true, deep synthesis. DeepRAG is not just an incremental RAG upgrade; it is a production-ready, agentic system engineered for enterprise-scale document intelligence. It enables agents to process and synthesize information across a corpus of up to 1,000 pages per query, delivering comprehensive, fully backed answers.

Why agents must evolve: The need for stateful reasoning

The failure mode of simple RAG lies in its statelessness and its susceptibility to the "retrieval bottleneck"—where generation quality is gated by the effectiveness of the initial single-shot document retrieval. To achieve true utility, enterprise agents must be designed to address complex data challenges:

• Conflict resolution: DeepRAG tracks evidence with source, timestamp, and author metadata to reason about contradictions and timeliness across documents.

• Auditability and compliance: It provides high-fidelity traceability for every finding, which is a non-negotiable requirement for regulated industries.

• Synthesize, not summarize: The system is optimized for cross-document reasoning to connect disparate facts and synthesize a cohesive answer from fragmented evidence.

How DeepRAG unlocks true comprehension

DeepRAG's power is rooted in a sophisticated, multi-step agentic reasoning workflow that mimics the process of a human expert conducting research. Instead of a single "retrieve-and-answer" step, DeepRAG operates in three distinct, stateful phases.

Phase 1: Initial planning and intent analysis

When a complex query is received, the agent first performs intent analysis to break the question down into a sequence of concrete sub-questions. This phase involves effort estimation and sets intelligent limits to constrain the research space, ensuring the process is goal-directed before any retrieval occurs.

Phase 2: Iterative research loop (plan-query-consolidate)

This is the core of DeepRAG's stateful knowledge construction. The agent executes its plan in a continuous cycle:

• Plan: Determine the next research step based on the current state of knowledge.

• Select tool/index: Choose the appropriate data source for the next search.

• Query and retrieve: Execute a targeted search against the context index.

• Extract and consolidate: Gather the relevant evidence and merge it with the existing knowledge state, revising the plan based on the newly acquired information.

Phase 3: Final synthesis and quality validation

Once the iterative loop is complete, all accumulated evidence is fed into the final generation step. This produces a single, coherent, and comprehensive answer. A final quality validation step integrates the sourcing information and checks the response for completeness and accuracy against the compiled evidence.

Technical use cases: Input, logic, and output templates

DeepRAG is already solving critical synthesis problems in production environments. Here are examples focusing on the technical data flow.

1. Healthcare: Medical record summarization (MRS)

Problem: Clinicians require a summary from a corpus of 20–400 pages of disparate patient documents (clinical notes, lab results, imaging reports). Technical input: PDF or TXT corpus of up to 1,000 pages per patient. DeepRAG logic: The agentic loop executes sub-queries for specific data points (e.g., find all current medications, identify all cardiac diagnoses). It then synthesizes findings to generate a structured output, with critical guidelines in the prompt to highlight conflicting information and provide source traceability for every clinical finding. Structured output: A multi-section summary including chief complaint and diagnoses, medical history (with onset dates), and current medications (with dosages and prescribing dates), all with detailed sourcing.

2. Financial services: Contract and covenant analysis

Problem: Analyzing commercial credit risk requires synthesizing a multi-file repository (master agreements, amendments, supporting schedules) to track variances and identify default provisions. Technical input: A repository of commercial credit agreements and associated documents. DeepRAG logic: The agent performs cross-document comparison and risk analysis. For tasks like closing disclosure review, it performs verification checks to compare original vs. final loan terms and validate compliance against standards like TRID. It is explicitly prompted to identify and track discrepancies. Structured output: A contract analysis summary listing key terms and covenants (including specific financial thresholds). It also flags discrepancies identified with an impact assessment and provides an approval status.

3. Pharmaceuticals: Tech transfer documentation

Problem: Transferring a manufacturing process requires consolidating and validating data across dozens of files—batch records, QC data, equipment specs, and regulatory submissions. The core requirement is proactive gap analysis to prevent quality issues. Technical input: A corpus of manufacturing and quality documentation. DeepRAG logic: The agent uses its synthesis capability to identify critical process parameters (CPPs) and in-process controls from multiple documents. Crucially, it performs a risk analysis to identify and assess gaps between the sending and receiving sites’ documentation. Structured output: A tech transfer summary detailing critical parameters (with ranges), quality specifications (with acceptance criteria), and a summary of gaps identified with an associated impact assessment.

Configuring DeepRAG: Implementation and prompt engineering

DeepRAG is deployed via Agent Builder and requires specific configurations to enable its advanced synthesis capabilities.

Prerequisites and index configuration

• Ingestion mode: DeepRAG requires "Advanced" ingestion mode for your context index, which enables the multi-document synthesis capability.

• Document specifications: Documents must be PDF or TXT format. The hard limits are 512 MB maximum file size per file and 1,000 pages per index/query. Citation support for TXT is on the roadmap.

• Document quality: Native PDFs are preferred for optimal text extraction. Scanned documents must be pre-processed with OCR. Cost for ingestion is calculated at 0.2 AIU per page.

Prompt engineering best practices

The quality of DeepRAG's synthesis is highly correlated with the structure and specificity of the prompt. A structured prompt is essential for demanding outputs:

• Structured template: Define the agent's role (e.g., medical professional, financial analyst), the explicit task, and strict requirements.

• Traceability enforcement: Use an explicit instruction: “Critical: For every finding, provide the source identifier.”

• Conflict handling: For dirty enterprise data, include instructions to identify the conflict explicitly, present all versions with their sources and timestamps, and recommend a resolution (e.g., prefer more recent information).

• Performance trade-off: DeepRAG prioritizes depth and quality, resulting in longer processing times (typically 2–5 minutes per query, but could be longer). For simple, instantaneous lookup, reserve the use of semantic search instead.

How to get started with DeepRAG today

You can enable this powerful comprehension engine for your agents directly within Agent Builder.

Ready to move your agents from simple retrieval to deep comprehension? Learn more here: DeepRAG release notes; DeepRAG how-to guide