Evidence-groundedintelligence workspace.
AverQel is a production backend for turning private knowledge, external sources, and governed tools into a research workflow that stays inspectable from upload to answer to action.
Research systems break at the seams.
The difficult part is not adding another chat box. It is making files, retrieval, model providers, live tools, permissions, and long-running work behave like one trustworthy system. AverQel addresses that seam with explicit ownership, evidence boundaries, and a runtime that exposes what is happening.
Ground first. Act second.
Documents become private, searchable evidence. Answers remain bounded by context. Agentic actions are explicit, observable, and approval-aware. That separation is the product: useful autonomy without turning the system into an ungoverned black box.
From source to decision.
INGEST
Documents enter through idempotent, tenant-scoped upload flows, pass a malware scan, and move into private object storage before background extraction and indexing.
GROUND
Extraction, OCR fallback, chunking, embeddings, hybrid retrieval, reranking, and citation mapping turn source material into defensible evidence.
REASON
Axiom answers from the available context, supports structured output, and keeps uncertainty visible instead of inventing unsupported claims.
ACT
DeepSpace can plan bounded productivity work, search the web, read URLs, use approved MCP tools, ask for clarification, and pause before side effects.
A system with boundaries.
A secure path from file to evidence
The backend handles PDF, text, Markdown, image OCR, DOCX, PPTX, XLSX, legacy Office conversion, and configured code/text formats. Uploads are validated, scanned by ClamAV, stored privately in MinIO, then processed by workers for extraction, chunking, embeddings, and retrieval.
Answers that stay attached to sources
The query layer classifies requests, routes them through retrieval and reranking, preserves citation metadata, and returns grounded responses. The system prompt explicitly tells Axiom to use only relevant context and to say when evidence is insufficient.
Productivity work with visible state
DeepSpace is a bounded productivity runtime with task planning, note and Library operations, web and URL tools, clarification pauses, cancellation, resumable runs, and streamed thinking/tool/status events. It is intentionally not an unrestricted shell agent.
Autonomy with a policy boundary
Connectors and MCP are conversation-scoped and tenant-owned. Read-only actions can run when policy allows them; writes, deletes, sends, and other external side effects pause for visible approval. OAuth credentials are encrypted and provider access is checked at execution time.
Safety is an architecture decision.
The backend treats security, observability, and recovery as part of the feature surface. These are the decisions that make AverQel credible beyond a demo.
Fail closed on required malware scanning: an unavailable scanner blocks document acceptance rather than silently reducing the security boundary.
Treat extracted documents as untrusted text: content is escaped before it can enter rendered notes, and original blobs stay private behind authenticated routes.
Keep tenant and user ownership in every document, conversation, memory, connector, and MCP path; cross-tenant access is tested as a security property.
Persist run checkpoints, ordered activity steps, cancellation flags, and approval decisions so a long task can be inspected, resumed, or stopped safely.
Block arbitrary shell, filesystem, and cURL access from the model-facing DeepSpace tool contract; capabilities are explicit, bounded, and auditable.
Instrument HTTP and database work with OpenTelemetry and expose health, metrics, provider, dashboard, and support routes for operational visibility.
Built to be checked.
Capability is never authority.
A provider may generate an answer, a connector may expose a tool, and an agent may propose an action. AverQel still keeps ownership, policy, evidence, approval, and audit state outside the model. That is the line between an assistant feature and an operating system for knowledge work.
AverQel is a platform study in grounded intelligence, governed autonomy, and production boundaries.