How junkz3/deepfield works
Deepfield turns uploaded document corpora into calibrated, multilingual specialist workspaces, then answers user questions with visually retrieved, page-exact evidence and confidence explanations. Its primary flow runs from corpus ingestion and workspace setup through agent planning, visual retrieval, tool operations, and cited verdicts or work orders; the 3D universe makes retrieval and citations visible. The README documents voice, translation, WebXR, and optional account controls. Several UI and integration details are unsampled, so those connections are shown only where documentation or excerpts support them.
Main entry points: loop.ts Agent Loop · App.tsx Application Shell · useStepRunner.ts Step Runner · DeepfieldStudio.tsx Studio Setup
Components 25
Workspace Setup
- Workspace State
src/web/store.tsx - Studio Setup
src/web/components/DeepfieldStudio.tsx - Workspace Calibration
src/agent/calibrate.ts - Specialist Team
src/agent/team.ts
Agent Runtime
- Agent Loop
src/agent/loop.ts - Model Driver
src/agent/driver.ts - Vultr Inference
src/vultr/client.ts - Workspace Tools
src/agent/tools.ts - Confidence Scoring
src/agent/confidence.ts
Knowledge Evidence
- Corpus Ingestion
src/web/ingest.ts - Taxonomy Retrieval
src/agent/taxonomy.ts
User Experience
- Application Shell
src/web/App.tsx - Step Runner
src/web/hooks/useStepRunner.ts - Conversation View
src/web/components/ConversationView.tsx - Work Order View
src/web/components/WorkOrderView.tsx - 3D Knowledge Universe
src/web/components/Galaxy3D.tsx - Page Viewer
src/web/components/PageLightbox.tsx - Page Translation
src/web/translate.ts - Voice Services
src/web/tts.ts
Deployment
- Speech Relay
tools/tts-relay/serve.py - Account Access
deploy/auth.mjs - Production Server
deploy/server.mjs
Other components
- Technician / User
- Inference Service
- Hosted Speech APIs
Connections 27
- Technician / User → Application Shell: uses
README.md - Application Shell → Workspace State: reads and updates
src/web/App.tsx - Technician / User → Corpus Ingestion: uploads corpus
README.md - Corpus Ingestion → Workspace State: adds documents
src/web/ingest.ts - Technician / User → Studio Setup: configures workspace
README.md - Studio Setup → Workspace Calibration: calibrates
src/web/components/DeepfieldStudio.tsx - Studio Setup → Specialist Team: sets up specialists
src/web/components/DeepfieldStudio.tsx - Application Shell → Step Runner: starts conversation steps
- Step Runner → Workspace State: reads and dispatches
src/web/hooks/useStepRunner.ts - Step Runner → Agent Loop: runs step
src/web/hooks/useStepRunner.ts - Step Runner → Voice Services: speaks verdict
src/web/hooks/useStepRunner.ts - Agent Loop → Model Driver: plans and retrieves
src/agent/loop.ts - Agent Loop → Taxonomy Retrieval: scopes candidates
src/agent/loop.ts - Agent Loop → Workspace Tools: selects operations
src/agent/loop.ts - Agent Loop → Confidence Scoring: scores verdict
src/agent/loop.ts - Agent Loop → Specialist Team: routes specialist
src/agent/loop.ts - Model Driver → Vultr Inference: uses live model driver
src/vultr/client.ts - Vultr Inference → Inference Service: requests inference
README.md - Conversation View → Step Runner: initiates step
src/web/components/ConversationView.tsx - Conversation View → Work Order View: shows work order
src/web/components/ConversationView.tsx - Application Shell → 3D Knowledge Universe: renders backdrop
src/web/App.tsx - Application Shell → Page Viewer: renders page viewer
src/web/App.tsx - Page Viewer → Page Translation: translates page
src/web/components/PageLightbox.tsx - Voice Services → Speech Relay: uses relay
README.md - Speech Relay → Hosted Speech APIs: calls hosted speech
README.md - Production Server → Account Access: supports accounts
deploy/server.mjs - Production Server → Inference Service: proxies requests
README.md