> ## Documentation Index
> Fetch the complete documentation index at: https://docs.knowledgestack.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Pipeline Routing

> The ingestion pipeline routes documents through different processing paths based on two inputs: **document type** and **ingestion mode**.

## Routing Parameters

| Parameter      | Values                                                     | Description                                          |
| -------------- | ---------------------------------------------------------- | ---------------------------------------------------- |
| Document type  | `PDF`, `DOCX`, `PLAINTEXT`, `IMAGE`, `XLSX`, `CSV`, `PPTX` | Detected from file extension at upload time          |
| Ingestion mode | `standard` (default), `high_accuracy`, `single_chunk`      | Controls conversion backend and chunking granularity |

These parameters are set when you call `POST /v1/documents/ingest` or `POST /v1/documents/{id}/ingest`.

### Smart Defaults

When you omit the ingestion mode, the system picks the best default for your file type:

| Document Type | Default Mode    |
| ------------- | --------------- |
| PDF           | `high_accuracy` |
| IMAGE, CSV    | `single_chunk`  |
| All others    | `standard`      |

## Routing Decision Tree

```
                   Ingestion starts
                          |
            +-------------v--------------+
            | ingestion_mode ==          |
            | high_accuracy?             |
            +------+---------------+-----+
                   | yes           | no
                   v               v
          High-Accuracy     +------------------+
          Path              | ingestion_mode   |
                            | == single_chunk? |
                            +---+----------+---+
                                | yes      | no
                                v          v
                        +----------+  +-----------+
                        | doc_type?|  | doc_type  |
                        +--+--+--+-+  | == XLSX?  |
                       /   |     \    +--+-----+--+
                 IMAGE  CSV/TXT  PDF     | yes | no
                   |      |       |      v     v
                 Simple  Text   PDF   XLSX  Standard
                 Image   Path   SC    Path  Path
                 Path          Path
```

## Processing Paths

### 1. High-Accuracy Path

**Best for:** PDFs and images where maximum extraction quality is needed.

**What happens:**

1. **Preparation** -- Download source, remove watermarks (PDF)
2. **High-accuracy conversion** -- Advanced extraction using a secondary engine that produces detailed content lists, images, and markdown
3. **Smart chunking** -- Walks through extracted content, builds section hierarchy, creates text/image/table chunks with token-based merging

**Chunking behavior:** Text items accumulate across heading boundaries until the token limit (default 512) is reached or a visual item (image/table) interrupts. Headings update section routing but do not force chunk boundaries.

### 2. Single-Chunk Path -- Simple Image

**Best for:** Standalone images (photos, screenshots, diagrams).

The uploaded image becomes a single IMAGE chunk directly. No preparation, conversion, or chunking is needed.

### 3. Single-Chunk Path -- Text-Based (CSV / Plain Text)

**Best for:** CSV files and plain text documents you want treated as a single unit.

1. **Preparation** -- Download source file
2. **Single chunk creation** -- CSV files become a single TABLE chunk (with a head+tail preview of the data); plain text files become a single TEXT chunk with full content

### 4. Single-Chunk Path -- PDF

**Best for:** Short PDFs you want treated as a single retrievable unit (e.g., a one-page form or receipt).

1. **Preparation** -- Extract page screenshots and text
2. **Single chunk creation** -- Creates one chunk from page screenshots and extracted text

### 5. Excel Path

**Best for:** `.xlsx` and `.xlsm` spreadsheet files.

1. **Preparation** -- Download source file
2. **Excel conversion** -- Parse workbook structure (sheets, cells, formulas, dependencies) into structured JSON
3. **Excel chunking** -- Create one section per sheet, with table and text chunks preserving spreadsheet structure

See [Excel Pipeline](/ingestion/excel-pipeline) for full details.

### 6. Standard Path (Default)

**Best for:** PDFs, DOCX, PPTX, and plain text under standard processing.

1. **Preparation** -- Download source, remove watermarks (PDF)
2. **Conversion** -- Vision-language model pipeline: OCR, table detection, image extraction, structured JSON output
3. **Chunking** -- Hybrid chunking with interleaved text, table, and image chunks in document order

## Common Tail Steps

After the format-specific processing, all paths converge to the same finishing steps:

| Step               | Purpose                                                       |
| ------------------ | ------------------------------------------------------------- |
| Vector cleanup     | Remove stale vectors from prior failed runs                   |
| Enrichment         | Generate LLM descriptions for images and summaries for tables |
| Embedding          | Generate vector embeddings and store in the vector database   |
| Statistics         | Calculate token counts, section and chunk statistics          |
| Version activation | For version upgrades: switch the active version               |
| Completion         | Mark the pipeline as completed                                |

## Validation Rules

These rules are checked before the workflow starts:

| Rule                                                      | Error |
| --------------------------------------------------------- | ----- |
| Explicit chunk type requires `single_chunk` mode          | 400   |
| Secondary taxonomy requires IMAGE chunk type              | 400   |
| Chunk type cannot be overridden for CSV/plain text        | 400   |
| Each ingestion mode only supports specific document types | 400   |
| Custom page DPI only applies to PDFs                      | 400   |

## Configuration

Key pipeline settings:

```yaml theme={null}
chunking:
  max_tokens: 512          # Token limit per chunk (all paths)
  overlap_tokens: 64       # Standard path only

high_accuracy:
  formula_enable: false
  table_enable: true
  document_timeout: 720    # Seconds for conversion
```