Ingestion Management System (IMS)

Document Ingestion Pipeline

PDF, Word, Excel and rich dossier exhibits are parsed, deduplicated, chunked, indexed, and vectorised entirely on practitioner hardware — zero bytes leave the sovereign enclave.

Case Workbench Engine Setup
6 Formats
Source Document Types
<1ms
Lexical Index Retrieval
High-Density
LexVector™ Projections
0 bytes
Cloud Data Leakage
Pipeline Architecture

7-Stage Ingestion Pipeline

Each document passes through a deterministic sequence from raw file to dual-indexed knowledge base entry.

Step 1
Drop & Detect
File format auto-detected; scanned PDFs rejected if <30 chars/page
Step 2
Hash & Dedup
SHA-256 hash check; renamed files auto-linked, skipping redundant processing
Step 3
Parse & Chunk
Deterministic section parsing into semantic chunks with balanced overlap
Step 4
Lexical Index
Local full-text index for sub-millisecond keyword and citation search
Step 5
Sovereign Vectors
LexVector™ (legal) or FinVector™ (financial) sovereign dense projections
Step 6
Fact Registry
Structured case facts extracted for real-time LSP autocomplete synchronization
Step 7
Register & Brief
Document registered; agent posts 3–5 bullet post-ingest brief in case chat

Supported Document Formats

FormatParser EngineVector CoreScanned FallbackStatus
PDF Multi-Layer Document Parser LexVector™ Visual OCR placeholder Live
DOCX / DOC Structured Document Engine LexVector™ Automated court layout renderer Live
XLSX / XLS / CSV Forensic Tabular Ingestion Core FinVector™ Visual Table OCR Live
Markdown (.md) Native Structured Parser LexVector™ N/A Live
Hypertext Notes (.wiki) Sovereign Dossier Parser LexVector™ N/A Live
Plain Text (.txt) Native UTF-8 Parser LexVector™ N/A Beta
Ingest Modes

Interactive vs Batch Ingestion

Configured per-case in workspace policy — choose supervision level based on the importance of the document set.

Interactive Mode

For critical documents — Resolution Plans, CIRP applications. The agent reads the document, posts a 3–5 bullet "what I found" brief, then pauses and asks which section to prioritise before committing vectors.

High supervision, guided emphasis
Focus on financial / §29A / §30(2) waterfall
Shapes chunk tagging & retrieval ranking

Batch Mode

For surveillance cases with large document volumes — financial reports, court registry extracts. Queues all ingestion jobs in an async task queue, processing sequentially to prevent SQLite lock collisions.

Async queue, zero DB lock collisions
Hash-dedup skips redundant OCR
ONNX idle eviction frees RAM after 5 min
Pipeline Features

Built-in Safety & Intelligence

Every stage of the pipeline is hardened for CIRP-grade reliability and legal privilege protection.

Async QueueNo Lock Collisions

Serialised Ingestion Queue

All concurrent ingestFile and registerUnprocessedFile tasks are serialised in an in-memory queue, awaiting vector indexing and KV extraction to complete before the next job starts. Eliminates SQLITE_BUSY errors under concurrent document uploads.

Zero DB deadlocks Sequential + safe
Hash TrackingAuto-Relink

Rename Sync & Deduplication

File system watcher debounces deletions by 200ms and matches incoming files using SHA-256 content hashes. Renamed documents are automatically relinked to existing SQLite records, bypassing OCR and vector re-processing entirely.

SHA-256 match Instant relink
Dual VectorContent-Routed

Dual-Domain Vector Routing

Document type is classified at ingestion by structured schema heuristics. Legal filings route into LexVector™ while financial schedules route into FinVector™. The domain index enforces strict statutory and forensic retrieval segregation during inference.

LexVector™ FinVector™
Scanned PDFEarly Reject

Scanned Document Guard

Fully scanned PDFs with average text density below 30 characters per page are rejected at upload with a clear warning. Hybrid pages with scanned tables inject a structured [!WARNING] placeholder block to flag unextracted visual data for the practitioner.

Rejects blank scans Hybrid warning
Neural RuntimeAuto-Eviction

Smart Model Idle Eviction

A background resource monitor tracks vector memory activity. Sovereign model pipelines are automatically evicted from unified memory after idle inactivity, releasing local RAM back to the operating system without manual intervention.

~250 MB freed Auto TTL
RelationalACID Safe

Ingestion Sequence Integrity

Evidential records are registered in the master case ledger with transaction consistency before lexical and vector index generation. This guarantees relational referential integrity across all evidential sections.

Register-first Constraint safe