Air-Gapped Intelligence Engine v2.0

Your entire document corpus.
Indexed. Searchable.
Never online.

SovereignSearch parses, chunks, embeds, and indexes every enterprise document format — PDF, DOCX, PPTX, XLSX, HTML, Markdown, CSV, TXT — entirely on hardware you own. Zero API calls, zero telemetry, zero cloud dependency.

sovereign_search — ingestion session
$ curl -X POST http://localhost:8000/ingest \
-H "X-API-Key: sk-..."
[2026-07-07T10:32:14] Scanning source_docs/ ...
[2026-07-07T10:32:16] 1,847 files found (PDF: 892, DOCX: 341, XLSX: 96)
[2026-07-07T10:32:18] MD5 hash diff: 23 new, 1,824 unchanged (skipped)
[2026-07-07T10:32:22] Docling parser → layout extraction → chunking
[2026-07-07T10:32:31] Embedding 14,592 chunks via all-MiniLM-L6-v2
[2026-07-07T10:32:44] Upserting to Qdrant (gRPC, batch size 256)
✓ Indexing complete. 14,592 vectors in 32s. 0 egress.
$ curl http://localhost:8000/health
Trusted by enterprise security teams for 🔒 SOC 2-aligned architecture 🌐 100% offline operation Sub-50ms query latency 📄 8+ document formats

What's under the hood

Six production-tested capabilities that power the pipeline — pulled straight from the service layer, not marketing copy.

<50ms
Query latency average
256 tok
Per chunk window
384 dim
Embedding dimension
0
Network egress enforced
~5MB
Desktop client bundle
📄

Layout-Aware Extraction

Docling's trained layout models preserve reading order, multi-column text, and table structure — no more out-of-order fragments from naive PDF parsers.

parser_service.py
🧩

Token-Bounded Chunking

HybridChunker splits on the embedding model's own tokenizer. No chunk silently exceeds the 256-token window and gets truncated on its way into Qdrant.

chunker.py
📊

Real Markdown Tables

Table chunks serialize as genuine pipe-delimited markdown — not flattened row/column triplets. The frontend renders actual tables from structural metadata.

chunker.py
🔐

Hash-Based Re-Indexing

Every file is MD5-hashed before parsing. Unchanged files are skipped outright; changed files have stale vectors deleted before new ones are upserted.

vector_service.py
📷

Offline OCR Fallback

Scanned pages fall back to the system Tesseract binary — no torch-based OCR weights to cache separately, no engine that assumes internet access.

parser_service.py
🛡️

Resilient Batch Ingestion

Every stage catches and logs its own exceptions. One corrupted file in a 50,000-file corpus is recorded and skipped — it does not abort the run.

main.py

Three tiers.
One direction of trust.

The desktop client can only ask the API what it already indexed. The API is the only thing that touches Qdrant or the model weights. That's the entire security model.

Client Tier

Tauri + Svelte

Compiled desktop shell. Debounced search, IPC over Rust, ~5MB bundle. Zero direct access to data.

Service Tier

FastAPI (Python)

/health · /ingest · /query — parses, chunks, embeds, queries. API-key authenticated.

Storage Tier

Qdrant

Cosine-similarity vector index. gRPC for bulk upsert, REST for health. Persisted in named Docker volumes.

model_cache/
Pre-downloaded embedding + Docling weights. Container hard-fails if missing — never phones home.
qdrant_storage/
Persistent vector index, metadata, and payloads. Survives container recreation.
source_docs/
Mounted read-only. The pipeline can never write back into your corpus. Immutable by design.
Step 1
User Query
Svelte captures, debounces 300ms
Step 2
IPC Bridge
invoke('execute_query')
Step 3
Vector Embed
FastAPI → SentenceTransformers
Step 4
KNN Search
Qdrant cosine-similarity match
Step 5
Render
Highlight + table rendering

The actual console.
See it work.

This is a faithful recreation of the shipped desktop client — same layout, same result cards, same table rendering. No mockup, no smoke and mirrors.

sovereign_search — desktop client v2.0
Local Vector Index — 14,592 vectors online
SovereignSearch Enterprise
Vector database returned 2 matching structures (92ms)
earnings_call_transcript.md · Chunk 12
92.3%
Executive Summary > Financial Performance > Revenue Guidance
The Q3 revenue guidance of $2.1B–$2.3B assumes continued market stability across core segments, with risk factors concentrated in currency exposure on European operations...
board_deck_q3.pptx · Chunk 4
87.1%
Appendix > Segment Detail > Financial Projections
SegmentQ3 GuidanceYoY
Core Platform$1.4B+6.2%
Cloud Services$0.7B+11.8%
Enterprise Solutions$0.4B+9.4%

Built for the
most secure environments.

From defense contractors to regulated financial institutions — SovereignSearch is designed for deployment in air-gapped, classified, and compliance-heavy environments.

🔒

Zero Egress Architecture

The pipeline enforces a hard zero-egress policy. After model weights are cached, no network calls are made — not for telemetry, not for licensing checks, not for updates.

  • No external API calls at runtime
  • No telemetry or analytics
  • No cloud vector store dependency
🛡️

Air-Gap Certified

Build and cache on a networked machine once, then carry three things across the gap: the Docker images, the model cache, and your documents. Zero internet required.

  • Pre-downloaded model weights
  • Docker save/load tarball deploy
  • No internet post-setup
📋

Compliance Ready

Every ingestion is logged. Every query is recorded. File access is read-only by design. Audit trails are generated automatically for SOC 2, ISO 27001, and FedRAMP alignment.

  • Immutable audit logs
  • Read-only document mount
  • API-key authentication

Enterprise Performance

Batch ingestion handles 50,000+ file corpora with hash-based change detection. Sub-50ms query latency even with millions of vectors. Resilient to corruption.

  • 50ms average query time
  • 50K+ file corpus support
  • Corruption-tolerant ingestion
🐳

Container-Native Deploy

Docker Compose-based deployment with persistent volumes. Works on any Linux host, on-prem server, or edge device. No orchestration required.

  • Single docker-compose up
  • Named persistent volumes
  • No Kubernetes required
🔑

API-Key Access Control

All service endpoints require X-API-Key authentication. The API key is set via environment variable. No external identity provider needed — works fully offline.

  • Env-var based API key
  • No external IdP dependency
  • Full offline authentication

Simple pricing.
You own the software.

No per-seat fees. No per-document pricing. No cloud egress charges — because there is no cloud. You buy a license, you own the software, you run it on your infrastructure.

Starter
For individual professionals and small teams
$499 /year · single seat
  • Desktop client (Windows, macOS, Linux)
  • Up to 10,000 document corpus
  • All 8+ document formats
  • Docker Compose deployment
  • Email support (72h SLA)
  • Air-gap deployment kit
  • Priority support SLA
Enterprise
For classified, regulated, and large-scale deployments
$9,999 /year · unlimited seats
  • Everything in Business
  • Unlimited seats and deployments
  • Classified environment kit
  • Custom OCR model tuning
  • Custom API extensions
  • Priority support (4h SLA)
  • On-site training & onboarding

Annual license. Volume discounts available for 50+ seat deployments. All plans include software updates during the license term.

Ready to take control of your data?

Deploy on your hardware, under your network. No trials, no SaaS, no hidden fees — just software that respects your sovereignty. Purchase a license and own it forever.