---
title: "Building a File Semantic Analyzer: Guarding Outbound Data at Scale with AI"
description: "Organizations struggled to determine whether files leaving their environment contained business-critical information, personal data, or non-essential content. Uber’s File Semantic Analyzer (FSA) addresses the limits of keyword- and pattern-based DLP by extracting text from documents, using OCR for image-based files, chunking large inputs, and sending context-rich text to a fine-tuned large language model. The model produces summaries, named entities, key phrases, inferred intent, classifications, and explanations, which a rule-based engine uses for policy enforcement while analysts validate high-fidelity findings and provide feedback. The system reduced incident-response time from multiple hours to a few minutes per case, saves an average of five minutes per file, and is expected to analyze 150,000 files annually; the post reports a 97% reduction in false positives."
---

# Building a File Semantic Analyzer: Guarding Outbound Data at Scale with AI

[Uber](https://yomu.fyi/company/uber) · Jun 4, 2026

**Type:** Problem & solution

## Summary

Organizations struggled to determine whether files leaving their environment contained business-critical information, personal data, or non-essential content. Uber’s File Semantic Analyzer (FSA) addresses the limits of keyword- and pattern-based DLP by extracting text from documents, using OCR for image-based files, chunking large inputs, and sending context-rich text to a fine-tuned large language model. The model produces summaries, named entities, key phrases, inferred intent, classifications, and explanations, which a rule-based engine uses for policy enforcement while analysts validate high-fidelity findings and provide feedback. The system reduced incident-response time from multiple hours to a few minutes per case, saves an average of five minutes per file, and is expected to analyze 150,000 files annually; the post reports a 97% reduction in false positives.

## Context

FSA is presented as a response to enterprise data egress through email attachments, cloud uploads, and USB drives. Traditional DLP relies on keywords, tokens, and pattern matching, which can generate false positives and miss sensitive documents whose wording is not obvious. The system targets files such as business documents, customer data, source code, and personal content.

## Approach / What changed

The system collects anonymized files labeled by subject-matter experts as Business Critical, Personal, or Neutral. It extracts text from PDFs, documents, spreadsheets, presentations, and HTML, applies OCR to image-based files, and uses context-preserving chunking for large inputs. A fine-tuned LLM summarizes content, extracts entities and key phrases, infers intent, and explains classifications. A rule-based engine enforces policies, while analysts validate findings and feed corrections back into the process.

## Takeaways

- FSA combines semantic classification with summaries, entity extraction, inferred intent, and explanations instead of relying only on keyword or pattern matching.
- Large files are processed through context-preserving chunking, while OCR converts text in image-based files before generative AI analysis.
- The reported results include a 97% reduction in false positives, incident-response times falling from hours to minutes, and an expected 150,000 files analyzed annually.

## FAQ

### How does FSA process large or image-based files?

It converts supported file formats into plain text and applies OCR to image-based files. Large inputs are split into semantically coherent chunks so token limits are less likely to hide information from the language model.

### How does the system reduce errors from LLM hallucinations?

Each classification includes a transparent explanation that analysts can compare with the file content. High-fidelity findings are also routed to human analysts for validation, and chunking helps prevent context loss in large documents.

### What does the rule-based layer do after AI analysis?

It uses summaries, extracted entities, inferred intent, and explanations to enforce policies. Examples include alerting on unreleased product specifications sent externally, or on customer PII in a marketing list emailed to a personal address.

**Tags:** [AI](https://yomu.fyi/topic/ai), [Data Classification](https://yomu.fyi/topic/data-classification), [Generative AI](https://yomu.fyi/topic/generative-ai), [Security](https://yomu.fyi/topic/security)

- Source: [Uber](https://www.uber.com/us/en/blog/ubers-file-analyzer)
- Source URL: https://www.uber.com/us/en/blog/ubers-file-analyzer
- Ingested by Yomu: 2026-09-15T18:14:22.678Z

[Read original post](https://www.uber.com/us/en/blog/ubers-file-analyzer)