Urjasoft - Software, AI & SaaS Engineering Urjasoft - Software, AI & SaaS Engineering
Prototype ID: LAB-07 Category: Enterprise AI Systems

Private Knowledge Search & Semantic Vector Index

Prototype testing local vector retrieval with query-time role-based access filtering.

A self-hosted semantic search prototype allowing enterprises to test dense vector retrieval across internal documentation without sending data to public cloud vector APIs.

The Architectural Problem

Regulated enterprises cannot send proprietary intellectual property or sensitive customer records to public cloud vector embedding endpoints.

Working Hypothesis

Self-hosted vector indices operating within private network boundaries can achieve sub-second semantic retrieval with strict SQL-level permission enforcement.

System Topology

Prototype Architecture & Data Pipeline

Multi-stage execution model separating probabilistic reasoning from deterministic state mutations.

1

01. Document Chunking

Source documents segmented with inherited user group security attributes.

RBAC Chunking
2

02. Vector Generation

Local embedding generation operating within private network boundary.

Local Embeddings
3

03. Secure Indexing

Vector indices stored alongside row-level security metadata in PostgreSQL.

Secure Store
4

04. Filtered Retrieval

Similarity search strictly pre-filtered by authenticated user security permissions.

Query Engine
Data Boundary Rule: No non-deterministic agent loop is permitted to execute writes directly to transactional databases without an intermediate policy gatekeeper.
Implementation Methodology

How We Engineered the Prototype

1. Document Chunking: Internal documents parsed and tagged with permission metadata.
2. Vector Generation: Embeddings generated using self-hosted model runners.
3. Secure Indexing: Vectors stored in local relational database with security tags.
4. Query Intersect: Similarity query intersected with user role permissions before return.
Current Working Capabilities
  • ✓ [Implemented] Self-hosted vector similarity search running inside local network boundary
  • ✓ [Implemented] Role-based access filters applied at SQL query execution time
  • ✓ [Planned] Automated connectors for remote document wikis and intranet repositories
Known Limitations & Unsolved Cases
  • × Embedding generation on local CPU hardware introduces indexing latency for large corpora
  • × Requires structured role-tagging during document ingestion
Engineering Takeaways & Architectural Findings

“Observed Finding: Enforcing security access filters directly within the database query prevents unauthorized document titles or snippets from leaking through similarity rankings.”

Experiment Parameters
Maturity Status
Prototype
Research Discipline
Enterprise AI Systems
Experiment Identifier
LAB-07
Tested Technologies
PHP 8.4 Laravel PostgreSQL Docker
Related Commercial Capability
API Development & Integrations

Findings from this research area inform our engineering capabilities in api development & integrations.

Explore Related Capability →
Need a custom proof of concept built for your company?
Request Technical Consultation
Additional Explorations

Related Lab Experiments

View All Initiatives →
Engineering Collaboration

Translating research into production software.

Our engineering team can evaluate your technical roadmap, suggest architecture patterns, and partner with you on high-assurance system execution.

Schedule Technical Scoping Call Explore Full Labs Catalog