Keeping Humans in the Loop

Gitana provides an enterprise workflow engine that lets you streamline and customize processes to quickly ingest, approve and publish your data -- all while keeping humans in the loop!

Human-in-the-Loop

Your APIs and AI-driven endpoints are only as good as the knowledge corpus that it works with. The fastest way to degrade answer quality, undermine trust, or expose your organization to risk is to let an unchecked firehose of content flow into your customer-facing endpoints.

The Gitana workflow engine puts a deliberate approval gate between your internal knowledge graph and the data sets that are delivered to your customer-facing RAG applications.

It allows you to review and approve new or changed content before it's indexed and embedded. Automation does the heavy lifting (extraction, chunking, validation, classification). Humans handle the ambiguous, high‑impact, or high‑risk decisions, with an auditable trail of who approved what and when.

Managing Risk

Most of the heavy lifting during the ingestion process is handled through automation. Automated processes include text extraction, data cleansing, OCR extraction, canonicalized formatting and more.

The Gitana ingestion process analyzes your incoming data to find metadata, entities, attributes, properties and relationships. The axiomatic reasoning engine works out how this data maps or modifies the existing ontology and automatically captures any data assertions for you.

The process further enhances your incoming data through the automated application of reasoning over your data to ensure data coherency, generate summmaries and apply classification, tags and taxonomies. The resulting data is labeled to sharpen its quality.

The result is a set of changes (either additions, modifications or deletes) that comprise a potential modification to your knowledge graph. Gitana stages these changes for you on an ingestion branch. Branches are Git-like workspaces that you keep track of the content changes (or deletas) in a space that is off the main line.

The workflow engine will now present these changes for human review along with an automated report that assesses the risk of the incoming set of changes. Automated checks look for personal information, toxicity, licensing violations, duplication and broken links.

A human reviewer receives notification of this new set of changes. The human reviewer looks over the changes and reviews the risk reports to determine how to route the changes. This may include one-click approval, rejection or a request for manual intervention/changes.

Regulatory Compliance and Auditing

Once approved, the incoming changes are merged into the main line or into a future release. The merges are done transactionally with a incrementing commit history so that you can, at any time, revert, quarantine and roll back commits. In the event of a need to do so, the blast radius is reduced by being able to see, commit-by-commit, where any errant data or contradiction came from.

The incoming knowledge retains information about how the generated changes were arrived at, who approved them and at what time. A full audit trail and version history is maintained that reveals how the content was modified and merged into the target.

This makes it straightforward to assert compliance with security, legal or regulatory policies (such as GDPR, HIPAA or SOC 2). The versioned knowledge graph, along with its audit trail and workflow history, provides a rich history that supports the assertions needed to satisfy regulatory compliance.

In the event that data must be removed, the knowledge graph and its relational capture of your information lets you quickly pivot on a property and find the places where it is referenced. This makes rapid depublication much easier and also fully verifiable.

The Bottom Line

Gitana's workflow engine lets you keep humans-in-the-loop. It is highly automated and streamlined when it needs to be. However, human review and approval is a necessary and essential step in these workflows. It provides a gateway to ensure that your knowledge graph is as accurate and consistent as possible.

When RAG-based applications fail, it's invariably because the corpus isn’t trustworthy.

A human‑in‑the‑loop approval engine turns ingestion from a blind pipeline into a governed, observable, and improvable process.