Source: http://www.poma-ai.com/docs/sdk/

# POMA AI SDK

`PrimeCut` is the recommended client in the `poma` package. Use it to ingest documents, collect typed results, store `.poma` archives, and feed structure-aware chunksets into your retrieval stack.

The package also exports `AsyncPrimeCut`, `PomaArchive`, `unpack`, and `generate_cheatsheets` for async workflows, archive handling, and retrieval post-processing.

The legacy `Poma` client still ships for compatibility. It is deprecated. New code should start with `PrimeCut` or `AsyncPrimeCut`.

> Looking for an older SDK snapshot? Use the version switch above this page.

## Start with the recommended path

```python
from poma import PrimeCut

client = PrimeCut()
result = client.ingest("example.pdf", show_progress=True)

print(len(result.chunksets))
print(result.chunksets[0].to_embed)
```

## Getting started

- [Installation](/sdk/getting-started/installation)
  Install `poma` and the integration extras you need.
- [Authentication](/sdk/getting-started/authentication)
  Create a `POMA_API_KEY` and configure your environment.
- [Quickstart](/sdk/getting-started/quickstart)
  Process your first document with `PrimeCut`.

## Core workflow

- Ingest a file with `PrimeCut.ingest(...)`
- Collect a previously submitted job with `PrimeCut.collect(...)`
- Work with `PomaResult`, `PomaChunk`, and `PomaChunkSet`
- Save or reopen `.poma` archives with `PomaArchive` or `unpack(...)`
- Generate structure-preserving cheatsheets with `generate_cheatsheets(...)`
- Connect the result to Qdrant, LangChain, or LlamaIndex

## Concepts

- [Ingestion](/sdk/concepts/ingestion)
- [Results and archives](/sdk/concepts/results-and-archives)
- [Cheatsheets](/sdk/concepts/cheatsheets)

## CLI

Prefer the terminal? The [`poma` CLI](/cli/) uses the same public API (v2) for ingest, job status, downloads, and basic account flows.

## MCP

Working from an agent? POMA ships two separate MCP servers — [`poma-mcp`](/mcp/poma-mcp) for PrimeCut document ingestion (`primecut_*` tools) and [`poma-grill-mcp`](/mcp/grill-mcp) for the Grill context engine (`grill_*` tools, including a hosted endpoint at `mcp.poma-ai.com`). Wire one or both into any MCP-aware client — Claude Code, Claude Desktop, Cursor — and you can call them directly from the agent. Start at the [MCP overview](/mcp/).

## Integrations and reference

- [Qdrant integration](/sdk/integrations/qdrant)
- [LangChain integration](/sdk/integrations/langchain)
- [LlamaIndex integration](/sdk/integrations/llamaindex)
- [PrimeCut reference](/sdk/reference/primecut)
- [AsyncPrimeCut reference](/sdk/reference/async-primecut)
- [Migration guide](/sdk/migration/poma-to-primecut)