01
Set up the Cohere Parse client
This implementation guide complements the Cohere Parse overview. Install the current Python SDK with pip install -U cohere, then create a cohere.ClientV2 using an API key stored outside committed source.
For production work, follow the live API reference. It currently supports document.type = image_url, including remote image URLs and data URIs. Cohere's quickstart demonstrates converting an image into a Base64 data URI before calling parse-v5.0.
02
Choose Markdown or blocks
Markdown is the default and is usually simplest for text chunking and retrieval. It can preserve text, lists and HTML-formatted tables, with image references and descriptions where available.
Use output_format="blocks" when the application needs ordered typed regions such as text, tables and images with location information. Blocks are useful for layout-aware retrieval, separate table handling and review interfaces that map content back to source regions.
03
Validate critical fields and preserve metadata
Store the original document ID, filename, page number, parse timestamp, model ID and output format alongside every result. That makes future re-parsing reproducible and prevents old and new extraction outputs from being mixed silently.
Cohere says Parse does not return confidence scores. Validate business-critical amounts, dates, account numbers, percentages, legal clauses and dense table rows before downstream automation acts on them.
04
Prepare Parse output for RAG
Do not destroy document structure immediately after extracting it. Prefer chunks that preserve natural boundaries such as a heading with its paragraphs, a complete table with nearby explanation, a form section or a page-specific image description.
A Cohere-centered retrieval path can use Parse for structure, Embed for retrieval representation and Rerank for ordering candidate evidence. Those products solve different stages; Parse is the ingestion layer rather than the final decision-maker.
05
Rate limits, cost and production checks
Cohere currently lists Parse at 500 requests per minute for both trial and production keys. Add concurrency control, retry handling, backoff for 429 responses, idempotent page IDs and monitoring instead of firing an uncontrolled batch at the documented ceiling.
The August 27 launch rate is $1.50 per 1,000 pages: 10,000 pages is $15 and 100,000 pages is $150 for Parse API page charges alone. Keep downstream storage, embeddings, reranking and LLM costs separate, and test the hardest real documents before production.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.