> ## Documentation Index
> Fetch the complete documentation index at: https://docs.dataharbor.co/llms.txt
> Use this file to discover all available pages before exploring further.
# Input Normalization
> See how DataHarbor normalizes JSON, CSV, YAML, and Markdown into one canonical data model before controls run.
# Input Normalization
**Different source formats. One controls pipeline.**
Before DataHarbor can apply privacy controls or transforms, it normalizes the upstream payload into a canonical JSON model.
```text theme={null}
Upstream payload → Normalize to canonical JSON → Apply controls → Format output → Respond
```
Normalization is the stage that makes one `controls` pipeline work across JSON, CSV, YAML, and Markdown. See [Data Pipeline](../core/data-pipeline) for the full request flow.
## Supported input formats
| Format | Content-Type | Normalized shape |
| -------- | ----------------------------------------------------- | --------------------------------------------------------- |
| JSON | `application/json`, `application/*+json` | Parsed directly |
| CSV | `text/csv` | Array of objects keyed by header columns |
| YAML | `text/yaml`, `application/yaml`, `application/x-yaml` | JSON-compatible mappings, sequences, and scalars |
| Markdown | `text/markdown` | Document object with `content` and optional `frontmatter` |
## When to set `input_format`
Most of the time, DataHarbor can infer the source format from the upstream `Content-Type` header.
Set `input_format` only when the upstream service returns a missing, generic, or incorrect content type.
```yaml theme={null}
version: "0.3"
input_format: markdown
objects:
_default:
controls:
- type: allow
fields: [frontmatter.title, content]
```
## CSV normalization
CSV input is parsed as headered rectangular data. The first row defines the column names, and each later row becomes an object.
```csv theme={null}
name,age,active
Alice,30,true
Bob,25,false
```
Normalizes to:
```json theme={null}
[
{ "name": "Alice", "age": 30, "active": true },
{ "name": "Bob", "age": 25, "active": false }
]
```
Type coercion is conservative: obvious booleans, integers, and decimals are converted. Everything else stays a string.
## YAML normalization
YAML input is normalized using the JSON-compatible subset of YAML.
```yaml theme={null}
name: Alice
tags:
- admin
- owner
active: true
```
Normalizes to:
```json theme={null}
{
"name": "Alice",
"tags": ["admin", "owner"],
"active": true
}
```
Advanced YAML constructs that do not map cleanly to JSON — including anchors, aliases, tags, merge keys, multi-document streams, and non-finite floats — are rejected with a descriptive error.
## Markdown normalization
Markdown input is handled in document mode. DataHarbor preserves the body as `content` and extracts YAML front matter into `frontmatter` when present.
```markdown theme={null}
---
title: API Guide
tags:
- auth
---
# API Guide
Authentication details here.
```
Normalizes to:
```json theme={null}
{
"frontmatter": {
"title": "API Guide",
"tags": ["auth"]
},
"content": "# API Guide\n\nAuthentication details here."
}
```
See [Markdown Input](../delivery/markdown) for full front matter rules, limitations, and detection nuances.
## Body sniffing
When the upstream `Content-Type` header is missing or unrecognized and no `input_format` is declared, DataHarbor inspects the response body to infer the format.
* Starts with `{` or `[` → JSON
* Two rows with matching comma-separated field counts → CSV
* Starts with `---` or a `key:` mapping → YAML
* Starts with valid front matter plus non-empty body, or starts with `#` → Markdown
If detection is inconclusive, DataHarbor defaults to JSON.
## Why normalization matters
* Controls always run against normalized JSON, not raw upstream bytes.
* The same field targeting rules work across source formats.
* Output formatting happens later, so you can normalize from one format and respond in another.
## Next steps
See where normalization fits in the request flow
Learn how the final governed payload is rendered
Dive into Markdown-specific rules and limitations