Find answers by meaning
Search over embedded text chunks with semantic similarity instead of relying only on exact keywords or filenames.
Vianordis--:--:-- UTCVianordis / ED. 02 / 2026
§ 07 — App support
Spectra
Sovereign document intelligence with extraction, semantic search, and cited AI answers.
Spectra turns PDFs, Office files, spreadsheets, and document collections into searchable, structured, AI-ready knowledge while keeping processing, storage, identity, and vector search inside a European-controlled architecture.
01 — Problem
Enterprise knowledge is trapped in documents that normal search cannot understand.
Contracts, policies, reports, spreadsheets, presentations, filings, and research documents contain critical facts, but teams often search manually, copy data into spreadsheets, or ask generic AI tools without source evidence. That creates slow review cycles, weak auditability, and avoidable data-residency risk.
- Keyword search misses meaning, context, table values, and related language across document sets
- Manual review of contracts, policies, evidence packs, and due-diligence files does not scale
- AI answers are hard to trust when they do not cite the exact source page or text segment
- Documents and embeddings are often sent to third-party clouds outside the desired jurisdiction
- PDFs, Word files, PowerPoint decks, spreadsheets, and CSV files require different processing paths
- Teams need both human document viewers and AI-native workflows over the same governed data
02 — Solution
Spectra converts document collections into governed, searchable knowledge.
Spectra combines a Next.js interface engine with a Python Cortex processing engine. Documents are stored in S3-compatible object storage, extracted into text chunks and tables, enriched with 1536-dimensional embeddings, indexed in PostgreSQL/pgvector, and made available for semantic search and cited AI chat.
- Ingest PDFs, Word documents, PowerPoint decks, Excel spreadsheets, and CSV files into tenant-scoped projects.
- Extract text, layout structure, tables, page numbers, and bounding-box source coordinates where the processor can preserve them.
- Generate embeddings through the AI Hub and search document chunks by semantic similarity using pgvector HNSW.
- Use chat sessions and referenced chunks so AI answers can be connected back to the source material.
03 — Benefits
Outcomes Spectra is designed to improve
Spectra reduces document review friction by combining extraction, evidence, semantic search, and AI workflows in one controlled platform.
Verify AI output
Connect responses to referenced chunks, page numbers, and bounding-box metadata so reviewers can inspect the source.
Extract structured evidence
Store extracted tables as JSON rows and headers alongside document metadata, page counts, word counts, and processing status.
Process mixed file collections
Route PDFs, Office documents, spreadsheets, and CSV files through specialized processors instead of one generic parser.
Keep data under control
Use tenant-scoped PostgreSQL records, S3-compatible storage, Keycloak identity, Infisical-managed secrets, and Kubernetes deployment.
Scale background processing
Use a PostgreSQL job queue with `FOR UPDATE SKIP LOCKED`, retries, stale-job recovery, and separate Cortex processing replicas.
04 — How it works
How Spectra works
- 01
Upload and organize documents
Create a tenant-scoped project, upload supported documents to S3-compatible storage, and register metadata such as filename, MIME type, size, processing status, and project context.
- 02
Extract, chunk, and enrich
Cortex downloads the document, selects the right processor, extracts text and tables, saves chunks with page and bounding-box metadata, then generates embeddings through AI Hub.
- 03
Search, ask, and verify
Users search semantically, ask questions through AI chat, inspect cited chunks in document viewers, and export or reuse structured evidence after review.
05 — Features
Core Spectra capabilities
Document projects
Group documents by tenant, team, use case, review package, investigation, client matter, or knowledge domain.
Multi-format ingestion
Support PDFs, Word documents, PowerPoint presentations, Excel spreadsheets, and CSV files through dedicated processing paths.
PDF extraction
Use the Cortex PDF processor with IBM Docling-oriented extraction for text, layout, tables, page data, and source-location metadata.
Office and spreadsheet processing
Use MarkItDown-oriented Office parsing plus pandas/openpyxl spreadsheet handling for decks, documents, workbooks, and CSV files.
Vector search
Store 1536-dimensional embeddings in PostgreSQL pgvector and use an HNSW cosine index for fast similarity search.
Structured table storage
Persist extracted tables with headers, rows, page number, table index, bounding box, chunk link, and metadata.
AI Hub integration
Generate embeddings and chat completions through the AI Hub with tenant headers, service name, request IDs, and streaming support.
Document viewers
Use PDF viewing with source highlights and spreadsheet viewing for human verification of extracted or cited content.
Processing jobs
Queue document processing, embedding generation, and table extraction jobs with attempts, priority, schedule, output, error state, and retry behavior.
GenUI chat sessions
Store chat sessions and messages with assistant actions, action data, referenced chunks, and token counts.
06 — Architecture
Architecture and security model
Spectra uses a twin-engine architecture: a Next.js interface for users and AI interactions, and a Python/FastAPI Cortex engine for document processing. PostgreSQL/pgvector, Ceph S3, Keycloak, AI Hub, Infisical, and Kubernetes provide the platform foundation.
- Next.js interface engine for authentication, UI, file handling, AI streaming, PDF viewer, spreadsheet viewer, GenUI component registry, and health endpoint.
- Python/FastAPI Cortex engine with health, readiness, metrics, structured JSON logging, API-key readiness checks, metrics bearer token checks, and background job runner.
- PostgreSQL spectra schema with projects, documents, chunks, extracted tables, jobs, chat sessions, and chat messages.
- pgvector extension with 1536-dimensional embeddings and HNSW cosine index for semantic retrieval.
- Ceph RGW S3-compatible object storage for original document files and signed upload/download access.
- AI Hub integration for embeddings, non-streaming chat completions, and streaming chat completions.
- Keycloak OIDC via NextAuth, requiring explicit tenant assignment from Keycloak groups.
- Kubernetes manifests for namespace, configmap, external secrets, interface deployment/service, cortex deployment/service, ingress, and kustomization.
07 — Use cases
Where Spectra fits
Legal and contract review
Search clauses, obligations, dates, risks, and contract changes across large document sets with source-backed answers.
Compliance and audit evidence
Extract policy requirements, map evidence, compare guidance, and create review packs with verifiable citations.
Financial due diligence
Analyze reports, filings, workbooks, covenants, portfolio documents, and deal-room evidence with structured extraction.
Research and knowledge management
Search papers, technical manuals, procedures, legacy documentation, and internal knowledge bases by meaning.
Healthcare and life sciences documentation
Review clinical, regulatory, protocol, trial, and scientific documents while preserving source traceability.
Operations document search
Make SOPs, incident reports, supplier documents, procurement packs, and internal process files searchable and AI-ready.
08 — Integrations
Integrations and platform connections
Spectra is designed to plug into the Vianordis/GSC platform stack for identity, storage, AI services, database search, secrets, observability, and Kubernetes operations.
Keycloak OIDC
Keeps access tied to organization accounts, roles, and sign-in policy.
NextAuth.js
Shows how this service fits into the wider Vianordis environment instead of standing alone.
PostgreSQL
Stores structured service records in a controlled tenant-aware backend.
pgvector
Shows how this service fits into the wider Vianordis environment instead of standing alone.
Ceph RGW S3-compatible storage
Provides the governed place where service data and files are stored.
AI Hub embeddings and chat completions
Shows how this service fits into the wider Vianordis environment instead of standing alone.
Infisical external secrets
Shows how this service fits into the wider Vianordis environment instead of standing alone.
Kubernetes
Shows how this service fits into the wider Vianordis environment instead of standing alone.
Istio ambient namespace mode
Shows how this service fits into the wider Vianordis environment instead of standing alone.
NGINX ingress and TLS
Shows how this service fits into the wider Vianordis environment instead of standing alone.
Prometheus metrics
Shows how this service fits into the wider Vianordis environment instead of standing alone.
IBM Docling-oriented PDF extraction
Shows how this service fits into the wider Vianordis environment instead of standing alone.
MarkItDown-oriented Office parsing
Shows how this service fits into the wider Vianordis environment instead of standing alone.
pandas/openpyxl spreadsheet processing
Shows how this service fits into the wider Vianordis environment instead of standing alone.
S3 signed URL workflows
Provides the governed place where service data and files are stored.
09 — Trust
Controls for sensitive document intelligence
Spectra is designed for documents that may contain contracts, customer data, financial records, policies, research, regulated evidence, or internal operational knowledge.
Tenant isolation
RLS policies on projects, documents, chunks, tables, jobs, sessions, and messages
Tenant context
Database policies use `app.current_tenant_id` session context
Identity
Keycloak OIDC with explicit tenant assignment through groups
Storage control
S3 keys include tenant and project context with Ceph RGW-backed storage
Source evidence
Page numbers, bounding boxes, referenced chunks, and extracted table metadata
Processing safety
Job attempts, retries, stale-job recovery, and failed-document status capture
Service auth
Cortex readiness and metrics endpoints protected by API key or bearer token
Observability
Structured JSON logging, Prometheus metrics, health checks, and readiness checks
Secrets
Database, S3, Keycloak, NextAuth, API, and metrics secrets via Infisical/Kubernetes
Deployment status
Implementation is marked ready for deployment; rollout and DNS validation remain deployment-scope tasks
10 — Pricing
Scope Spectra by document volume, processing depth, and deployment model
Spectra pricing should reflect number of users, tenants, projects, documents, storage volume, page volume, embedding/chat usage, processing concurrency, retention requirements, and whether the deployment is managed, dedicated, or customer-controlled.
11 — FAQ
Questions technical buyers usually ask.
Who is Spectra for?
Spectra is for legal, compliance, finance, research, operations, healthcare, life-sciences, and knowledge teams that need controlled document extraction and semantic search.
Which file formats does Spectra support?
The implementation targets PDF, DOCX, DOC, PPTX, PPT, XLSX, XLS, and CSV processing through PDF, Office, and spreadsheet processors.
How does semantic search work?
Spectra chunks extracted text, generates 1536-dimensional embeddings through AI Hub, stores them in PostgreSQL pgvector, and searches by cosine similarity.
Can Spectra cite the source of an AI answer?
Yes. The data model stores referenced chunks, page numbers, and bounding-box metadata so answers can be tied back to document locations where available.
Does Spectra extract tables?
Yes. The schema stores extracted tables with headers, rows, page number, table index, bounding box, and metadata, and Cortex includes table extraction job support.
Where are source documents stored?
Documents are stored in S3-compatible object storage. The inspected configuration uses Ceph RGW with the `spectra-documents` bucket.
How is tenant isolation handled?
Spectra uses tenant IDs across projects, documents, chunks, tables, jobs, chat sessions, and messages, with PostgreSQL row-level security policies tied to tenant session context.
Does Spectra include an AI chat interface?
Yes. The interface includes AI Hub chat completion helpers, streaming support, chat session/message tables, and GenUI action/component plumbing.
Is Spectra already deployed?
The inspected status document marks the implementation and images as ready for deployment, while registry push, Kubernetes deployment, DNS setup, workflow testing, and AI Hub verification are listed as remaining rollout steps.
Can Spectra be self-hosted?
Yes. The app includes Kubernetes manifests for the interface and Cortex services, external secrets, ingress, config, namespace, and kustomization.
12 — Next step
Turn document collections into searchable, source-backed intelligence.
Use Spectra to ingest files, extract text and tables, generate embeddings, ask cited questions, and keep document intelligence inside a governed sovereign platform.