# Architecture

This chapter describes how ELS Compass is built and how a request travels
through it, from an uploaded file to a cited answer.

The material comes from the backend engineering documentation and tracks the
implementation as it stands. Where a page and the running code disagree, the
code wins and the page needs updating.

| Page | What it covers |
| --- | --- |
| [Overview](overview.md) | Layered structure, composition root, storage split, startup behaviour, request flow diagram. |
| [Request routing](request-routing.md) | How an incoming question is classified and routed. |
| [Document ingestion](document-ingestion.md) | The ingestion pipeline stage by stage, the durable job runner, and the catalog and retrieval-pyramid projections. |
| [Chunking](chunking.md) | Chunk layers, section ordinal metadata, extractive and abstractive summaries. |
| [Answering](answering.md) | Conversation orchestration, Creative and Grounded evidence regimes, workload planning, retrieval, reranking, answer publication. |
| [Model coordination](model-coordination.md) | Shared model placement, leases and residency state across processes and nodes. |
| [Document lifecycle](document-lifecycle.md) | Document states, allowed transitions, reprocessing rules. |
| [Workspace and conversation governance](workspace-conversation-governance.md) | Workspace lifecycle, conversation boundary, archive and delete rules. |
| [Design decisions](design-decisions.md) | Implementation choices that matter today, plus deliberate deferrals. |

```{toctree}
:maxdepth: 2
:hidden:

overview
request-routing
document-ingestion
chunking
answering
model-coordination
document-lifecycle
workspace-conversation-governance
design-decisions
```