> ## Documentation Index
> Fetch the complete documentation index at: https://docs.enterspeed.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Content sources
> Indexes, stage outputs, and how a task's input is chosen at session time rather than baked into the configuration.
A content source is the data a task runs over. Speedtrain has two kinds.
## Indexes
An **index** holds documents ingested from an external system — a PIM, ERP, CMS, or anything else that can push JSON. Each document is a record: one product, one article, one row.
Speedtrain reads indexes; it never writes back to them. Documents are managed by whoever pushes them in.
Find them under **Content sources → Indices**.
## Stage outputs
A **stage output** is the content one task configuration has produced. Feeding it into a second task is how you chain jobs:
```
index ──▶ task 1: extract attributes ──▶ stage output ──▶ task 2: write copy
```
Find them under **Content sources → Stage outputs**.
A chained stage carries no images. When the input is another task's output there is no source document and therefore no `$images` — so a configuration that needs to see pictures must run first, directly on the index.
## The input is chosen at session time
This is the design decision worth internalising: **a task configuration does not name its input.**
The input is bound when a review session starts. One configuration can therefore run over different indexes, over a filtered subset of an index, or over another task's output, with no change to the configuration itself.
When starting a session you choose between:
A raw index, optionally narrowed by a selection predicate so the session covers only part of it.
Another task's output. By default this means that task's **published** version, accepted items only.
The exact outputs of one specific past run. Useful for pipeline testing, because stage two then consumes precisely what a chosen stage-one session accepted.
Testing sessions get the richer options — selection predicates, version selectors, filters. Real review sessions are deliberately stricter: they identify the feed and nothing else, so what runs in production is unambiguous.
## Document shape
A document is whatever JSON you push. Use your source system's own field names.
One key is reserved: a top-level `$images` array attaches pictures to the record, so that a vision-capable model can see them. [Add a content source](/speedtrain/getting-started/content-source)