Grooper and AI - Revision history

Dgreenwood: /* Azure LLM Quickstart */

2026-04-14T16:00:52Z

Azure LLM Quickstart

← Older revision		Revision as of 11:00, 14 April 2026
Line 338:		Line 338:
	# In the "URL" property, enter the "Target URI" from Azure		# In the "URL" property, enter the "Target URI" from Azure
	#:*<li class="fyi-bullet"> For Azure OpenAI model deployments this will resemble:		#:*<li class="fyi-bullet"> For Azure OpenAI model deployments this will resemble:
	#:**<code><nowiki>https://{your-resource-name}.openai.azure.com/openai/deployments/{~~model~~-id}/chat/completions?api-version={api-version}</nowiki></code> for Chat Service deployments		#:**<code><nowiki>https://{your-resource-name}.openai.azure.com/openai/deployments/{deployment-name}/chat/completions?api-version={api-version}</nowiki></code> for Chat Service deployments
	#:**<code><nowiki>https://{your-resource-name}.openai.azure.com/openai/deployments/{~~model~~-id}/embeddings?api-version={api-version}</nowiki></code> for Embeddings Service deployments		#:***<li class="attn-bullet"> At this time, the URL '''must''' take this format. If you copy a URL wth a <code>responses</code> endpoint from Azure, it will not work in Grooper. You will need to manually create the URL, filling in your resource's name and deployment name. Be sure to use the deployment's name and '''not''' the model's name.
			#:**<code><nowiki>https://{your-resource-name}.openai.azure.com/openai/deployments/{deployment-name}/embeddings?api-version={api-version}</nowiki></code> for Embeddings Service deployments
	#:*<li class="fyi-bullet"> For other models deployed in Azure AI Foundry (formerly Azure AI Studio) this will resemble:		#:*<li class="fyi-bullet"> For other models deployed in Azure AI Foundry (formerly Azure AI Studio) this will resemble:
	#:** <code><nowiki>https://{model-id}.{your-region}.models.ai.azure.com/v1/chat/completions</nowiki></code> for Chat Service deployments		#:** <code><nowiki>https://{model-id}.{your-region}.models.ai.azure.com/v1/chat/completions</nowiki></code> for Chat Service deployments
Line 392:		Line 393:
	* [https://learn.microsoft.com/en-us/azure/ai-studio/ Azure AI Studio full documentation]		* [https://learn.microsoft.com/en-us/azure/ai-studio/ Azure AI Studio full documentation]
	* [https://azure.microsoft.com/en-us/products/ai-model-catalog Azure AI model catalog]		* [https://azure.microsoft.com/en-us/products/ai-model-catalog Azure AI model catalog]


	== OpenAI and Azure OpenAI rate limits ==		== OpenAI and Azure OpenAI rate limits ==

Dgreenwood: /* Grooper before AI and after AI */

2026-04-09T13:38:31Z

Grooper before AI and after AI

@@ Line 577: / Line 577: @@
 '''Result:''' ''Days of image tuning reduced to zero configuration in most deployments.''
 |}
 === Document separation ===
@@ Line 620: / Line 621: @@
 * New Document Types meant retraining from scratch
 * Edge cases and lookalike documents caused misclassification
 | style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" |
 Classify documents by describing them in natural language — no samples or training required.
@@ Line 658: / Line 658: @@
 | colspan="2" style="background:#f0f8f5; padding:10px; color:#0f6e56;" |
 '''Result:''' ''A single natural language prompt replaces dozens of fragile field rules and vendor templates.''
 |}

Dgreenwood: /* Grooper before AI and after AI */

2026-04-08T17:08:04Z

Grooper before AI and after AI

Dgreenwood: /* Grooper before AI and after AI */

2026-04-08T16:10:42Z

Grooper before AI and after AI

← Older revision		Revision as of 11:10, 8 April 2026
Line 531:		Line 531:

	<big>This used to be hard</big>		<big>This used to be hard</big>

	''Less configuration. Better results.''		''Less configuration. Better results.''

Dgreenwood at 15:59, 8 April 2026

2026-04-08T15:59:51Z

← Older revision		Revision as of 10:59, 8 April 2026
Line 527:		Line 527:
	Refer to the documentation for your chosen model to verify supported parameters before modifying default settings.		Refer to the documentation for your chosen model to verify supported parameters before modifying default settings.
	\|}		\|}

			== Grooper before AI and after AI ==

			<big>This used to be hard</big>

			''Less configuration. Better results.''

			AI has fundamentally changed how the Grooper platform is configured — leading to accelerated deployments and systems that are easier to manage.

			{\| class="wikitable" style="width:100%"
			\|-
			! style="width:50%; background:#eeeeee;" \| The old way
			! style="width:50%; background:#e1f5ee; color:#085041;" \| The new way
			\|-
			\| style="vertical-align:top; padding:10px;" \|
			* Relied on lots of manual configuration
			* Tedious and time-consuming
			* Difficult to manage at scale
			* Required deep platform knowledge up front
			\| style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" \|
			* AI-driven and largely automatic
			* Configured with natural language prompts
			* Easier to troubleshoot and maintain
			* Accessible without years of platform expertise
			\|}

			=== Image cleanup for OCR ===

			''Defects on images interfere with traditional OCR engines, leading to poor accuracy and data capture failures.''

			{\| class="wikitable" style="width:100%"
			\|-
			! style="width:50%; background:#eeeeee;" \| Old way — IP Profiles
			! style="width:50%; background:#e1f5ee; color:#085041;" \| New way — Azure DI OCR
			\|-
			\| style="vertical-align:top; padding:10px;" \|
			Build Image Processing Profiles to adjust the image prior to running an OCR engine.

			* No universal set of commands works across all image types and sources
			* Extensive configuration and testing required per job
			* Complexity compounds with every step added to a profile
			\| style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" \|
			An AI-based OCR engine trained on millions of images of varied types and quality.

			* AI engine requires no pre-OCR cleanup in nearly all cases
			* Works out of the box with zero image tuning
			* Handles a wide variety of image qualities automatically
			\|-
			\| colspan="2" style="background:#f0f8f5; padding:10px; color:#0f6e56;" \|
			'''Result:''' ''Days of image tuning reduced to zero configuration in most deployments.''
			\|}
			=== Document separation ===

			''Scanned or digital page collections need to be split into individual documents — one of the hardest problems in document processing.''

			{\| class="wikitable" style="width:100%"
			\|-
			! style="width:50%; background:#eeeeee;" \| Old way — 7 Separation Providers
			! style="width:50%; background:#e1f5ee; color:#085041;" \| New way — AI Separate
			\|-
			\| style="vertical-align:top; padding:10px;" \|
			Choose from and configure one of seven different Separation Providers.

			* Required expertise in each provider's logic and limitations
			* Each provider designed for narrow, predictable separation scenarios
			* Most providers required heavy configuration
			* ESP Auto Separation added curated training data complexity
			\| style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" \|
			A Separation Provider that uses an LLM to identify document boundaries from natural language instructions.

			* Instructions written in plain, human-readable language
			* Handles unpredictable separation points with ease
			* Optional reasoning output makes troubleshooting simple
			\|-
			\| colspan="2" style="background:#f0f8f5; padding:10px; color:#0f6e56;" \|
			'''Result:''' ''Separation logic that once required training data and specialist knowledge is now written in plain English.''
			\|}

			=== Document classification ===

			''Documents of different types need to be identified and routed correctly before data can be extracted from them.''

			{\| class="wikitable" style="width:100%"
			\|-
			! style="width:50%; background:#eeeeee;" \| Old way — Training-based classification
			! style="width:50%; background:#e1f5ee; color:#085041;" \| New way — AI Classify
			\|-
			\| style="vertical-align:top; padding:10px;" \|
			Curate sample documents for each document type and train a classification model.

			* Required curated sample documents per document type
			* Training sets needed regular maintenance as formats evolved
			* New document types meant retraining from scratch
			* Edge cases and lookalike documents caused misclassification
			\| style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" \|
			Classify documents by describing them in natural language — no samples or training required.

			* No sample documents or training required to get started
			* Handles novel document types immediately
			* Reasoning output explains every classification decision
			\|-
			\| colspan="2" style="background:#f0f8f5; padding:10px; color:#0f6e56;" \|
			'''Result:''' ''Adding a new document type goes from a multi-day training exercise to writing a paragraph.''
			\|}

			=== Data extraction ===

			''Specific fields need to be located and extracted from documents — even when layouts vary across vendors, versions, or sources.''

			{\| class="wikitable" style="width:100%"
			\|-
			! style="width:50%; background:#eeeeee;" \| Old way — Field rules and templates
			! style="width:50%; background:#e1f5ee; color:#085041;" \| New way — AI Extract
			\|-
			\| style="vertical-align:top; padding:10px;" \|
			Build zone-based or rule-based extraction definitions for each document type and vendor.

			* Zone-based extraction broke whenever layouts changed
			* Regex and keyword rules required technical expertise
			* Each vendor or form variant needed its own template
			* Maintenance burden grew with every new document source
			\| style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" \|
			Describe what to extract in plain language. Works across varied formats automatically.

			* Layout-agnostic — works across varied formats without template changes
			* One extraction definition handles multiple vendors
			* Contextual understanding catches fields rule-based tools miss
			\|-
			\| colspan="2" style="background:#f0f8f5; padding:10px; color:#0f6e56;" \|
			'''Result:''' ''A single natural language prompt replaces dozens of fragile field rules and vendor templates.''
			\|}

			----

Dgreenwood: /* LLM connectivity and constructs */

2026-03-18T21:05:02Z

LLM connectivity and constructs

← Older revision		Revision as of 16:05, 18 March 2026
Line 39:		Line 39:

	=== LLM connectivity and constructs ===		=== LLM connectivity and constructs ===
	~~<section begin="llm-constructs" />~~
	====LLM Connectivity====		====LLM Connectivity====

Line 46:		Line 46:
	:** <li class="fyi-bullet"> Compatible APIs must support <code>chat/completions</code> and <code>embeddings</code> endpoints similar to the OpenAI API to interoperate with Grooper’s LLM features.		:** <li class="fyi-bullet"> Compatible APIs must support <code>chat/completions</code> and <code>embeddings</code> endpoints similar to the OpenAI API to interoperate with Grooper’s LLM features.
	:* '''Azure''' – Connects Grooper to LLMs offered by Microsoft Azure through its Model Catalog, including Azure OpenAI models.		:* '''Azure''' – Connects Grooper to LLMs offered by Microsoft Azure through its Model Catalog, including Azure OpenAI models.
			<section begin="llm-constructs" />
	====LLM-enabled extraction capabilities====		====LLM-enabled extraction capabilities====

Dgreenwood: /* LLM connectivity and constructs */

2026-03-18T21:04:35Z

LLM connectivity and constructs

← Older revision		Revision as of 16:04, 18 March 2026
Line 39:		Line 39:

	=== LLM connectivity and constructs ===		=== LLM connectivity and constructs ===
			<section begin="llm-constructs" />
	====LLM Connectivity====		====LLM Connectivity====

Line 94:		Line 94:

	:[[Prompt Engineering]] – The practice of designing and refining prompts to obtain desired responses from an LLM.		:[[Prompt Engineering]] – The practice of designing and refining prompts to obtain desired responses from an LLM.
			<section end="llm-constructs" />

	=== AI Search and the Search page ===		=== AI Search and the Search page ===

Dgreenwood: /* Known issues with certain models */

2026-02-04T16:41:00Z

Known issues with certain models

← Older revision		Revision as of 11:41, 4 February 2026
Line 501:		Line 501:
	=== Known issues with certain models ===		=== Known issues with certain models ===

	Occasionally, newly released models may not behave the same as previous models when used in Grooper. When this occurs, the table below will be updated with known issues, their impact on Grooper functionality, and any recommended workarounds.		Occasionally, newly released models may not behave the same as previous models when used in Grooper. Or, Parameters configurations may not be supported. When this occurs, the table below will be updated with known issues, their impact on Grooper functionality, and any recommended workarounds.

	{\| class="wikitable" style="margin:auto"		{\| class="wikitable" style="margin:auto"
Line 513:		Line 513:
	\|-		\|-
	\| gpt-5.1 (and above) \|\| 1/30/2026 \|\| Grooper does not currently expose a ''none'' option for the '''Reasoning Effort''' parameter. Development is evaluating how to address this best but has not implemented a solution at the time this is written.		\| gpt-5.1 (and above) \|\| 1/30/2026 \|\| Grooper does not currently expose a ''none'' option for the '''Reasoning Effort''' parameter. Development is evaluating how to address this best but has not implemented a solution at the time this is written.
			\|-
			\| Structured Outputs supported models \|\| NA \|\| OpenAI's Structured Outputs feature is designed to make the model stick to a developer-supplied JSON schema. Paradoxically, there are scenarios where Structured Outputs can cause the model to be thrown in a repetitive loop when creating its JSON response. If you find this issue is affecting you, change the "Use Structured Output" property to ''False''.
	\|}		\|}

Dgreenwood: /* Known issues with certain models */

2026-02-04T16:31:01Z

Known issues with certain models

← Older revision		Revision as of 11:31, 4 February 2026
Line 507:		Line 507:
	! Model name !! Date reported !! Notes		! Model name !! Date reported !! Notes
	\|-		\|-
	\| gpt-5 \|\| ~~1/30/2026~~ \|\| gpt-5 does not support the use of the '''Temperature''' ~~parameter when configuring model~~ parameters. Use the default Temperature ~~setting~~ to avoid a Bad Request error.		\| gpt-5 (and above) \|\| NA \|\| gpt-5 does not support the use of the '''Temperature''' or '''Top P''' parameters. Use the default Temperature and Top P settings in Grooper to avoid a Bad Request error.
			:*<li class="fyi-bullet"> gpt-5.2 does support these parameters when reasoning effort is set to ''none''. However, Grooper does expose a ''none'' option at this time.
			\|-
			\| gpt-5.1 (and above) \|\| NA \|\| gpt-5.1 does not support the ''minimal'' option for the '''Reasoning Effort''' parameter. Use ''low'', ''medium'' or ''high'' to avoid a Bad Request error.
			\|-
			\| gpt-5.1 (and above) \|\| 1/30/2026 \|\| Grooper does not currently expose a ''none'' option for the '''Reasoning Effort''' parameter. Development is evaluating how to address this best but has not implemented a solution at the time this is written.
	\|}		\|}

Dgreenwood: /* OpenAI and Azure OpenAI rate limits */

2026-02-04T15:02:01Z

OpenAI and Azure OpenAI rate limits

Show changes

← Older revision		Revision as of 12:08, 8 April 2026
Line 537:		Line 537:
	{\| class="wikitable" style="width:100%"		{\| class="wikitable" style="width:100%"
	\|-		\|-
	! style="width:50%; background:#eeeeee;" \| ~~The old~~ way		! style="width:50%; background:#eeeeee;" \| Old way
	! style="width:50%; background:#e1f5ee; color:#085041;" \| ~~The new~~ way		! style="width:50%; background:#e1f5ee; color:#085041;" \| New way
	\|-		\|-
	\| style="vertical-align:top; padding:10px;" \|		\| style="vertical-align:top; padding:10px;" \|
Line 562:		Line 562:
	\|-		\|-
	\| style="vertical-align:top; padding:10px;" \|		\| style="vertical-align:top; padding:10px;" \|
	Build Image Processing Profiles to adjust the image prior to running an OCR engine.		Build IP (Image Processing) Profiles to adjust the image prior to running an OCR engine.

	* No universal set of commands works across all image types and sources		* No universal set of commands works across all image types and sources
Line 568:		Line 568:
	* Complexity compounds with every step added to a profile		* Complexity compounds with every step added to a profile
	\| style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" \|		\| style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" \|
	An AI-based OCR engine trained on millions of images of varied types and quality.		Uses Azure Document Intelligence to analyze images, providing an AI-based OCR engine trained on millions of images of varied types and quality.

	* ~~AI engine requires~~ no pre-OCR cleanup in nearly all cases		* Requires no pre-OCR cleanup in nearly all cases
	* Works out of the box with zero image tuning		* Works out of the box with zero image tuning
	* Handles a wide variety of image qualities automatically		* Handles a wide variety of image qualities automatically
Line 579:		Line 579:
	=== Document separation ===		=== Document separation ===

	''Scanned or digital page collections need to be split into individual documents — one of the hardest problems in document processing.''		''Scanned or digital page collections need to be split into individual documents — often one of the hardest problems in document processing.''

	{\| class="wikitable" style="width:100%"		{\| class="wikitable" style="width:100%"
Line 592:		Line 592:
	* Each provider designed for narrow, predictable separation scenarios		* Each provider designed for narrow, predictable separation scenarios
	* Most providers required heavy configuration		* Most providers required heavy configuration
	* ESP Auto Separation ~~added~~ curated training data complexity		* Using "ESP Auto Separation" adds curated training data complexity
	\| style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" \|		\| style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" \|
	A Separation Provider that uses an LLM to identify document boundaries from natural language instructions.		A Separation Provider that uses an LLM to identify document boundaries from natural language instructions.
Line 610:		Line 610:
	{\| class="wikitable" style="width:100%"		{\| class="wikitable" style="width:100%"
	\|-		\|-
	! style="width:50%; background:#eeeeee;" \| Old way — ~~Training-based classification~~		! style="width:50%; background:#eeeeee;" \| Old way — Lexical Classification
	! style="width:50%; background:#e1f5ee; color:#085041;" \| New way — ~~AI Classify~~		! style="width:50%; background:#e1f5ee; color:#085041;" \| New way — LLM Classifier
	\|-		\|-
	\| style="vertical-align:top; padding:10px;" \|		\| style="vertical-align:top; padding:10px;" \|
	Curate sample documents for each ~~document type~~ and train a classification model.		Curate sample documents for each Document Type and train a classification model.

	* Required curated sample documents per ~~document type~~		* Required curated sample documents per Document Type
	* Training sets needed regular maintenance as formats evolved		* Training sets needed regular maintenance as formats evolved
	* New ~~document types~~ meant retraining from scratch		* New Document Types meant retraining from scratch
	* Edge cases and lookalike documents caused misclassification		* Edge cases and lookalike documents caused misclassification

	\| style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" \|		\| style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" \|
	Classify documents by describing them in natural language — no samples or training required.		Classify documents by describing them in natural language — no samples or training required.
Line 628:		Line 629:
	\|-		\|-
	\| colspan="2" style="background:#f0f8f5; padding:10px; color:#0f6e56;" \|		\| colspan="2" style="background:#f0f8f5; padding:10px; color:#0f6e56;" \|
	'''Result:''' ''Adding a new ~~document type~~ goes from a multi-day training exercise to writing a paragraph.''		'''Result:''' ''Adding a new Document Type goes from a multi-day training exercise to writing a paragraph.''
	\|}		\|}

Line 637:		Line 638:
	{\| class="wikitable" style="width:100%"		{\| class="wikitable" style="width:100%"
	\|-		\|-
	! style="width:50%; background:#eeeeee;" \| Old way — ~~Field rules~~ and templates		! style="width:50%; background:#eeeeee;" \| Old way — Extractors and templates
	! style="width:50%; background:#e1f5ee; color:#085041;" \| New way — AI Extract		! style="width:50%; background:#e1f5ee; color:#085041;" \| New way — AI Extract
	\|-		\|-
	\| style="vertical-align:top; padding:10px;" \|		\| style="vertical-align:top; padding:10px;" \|
	Build ~~zone~~-based or ~~rule~~-based ~~extraction definitions~~ for each document ~~type and vendor~~.		Build rule-based (or worse zone-based) extractors for each field, table and section for each document.

	* Zone-based ~~extraction~~ broke whenever layouts changed		* Zone-based extractors broke whenever layouts changed
	* ~~Regex and keyword rules~~ required technical expertise		* Rules-based extractors use regex or required other technical expertise
	* Each vendor or form variant ~~needed~~ its own template		* Each vendor or form variant may need its own template (Document Type) and override logic
			* Too much to learn: 25 Value Extractors. 7 Table Extract Methods. 11 Section Extract Methods.
	* Maintenance burden grew with every new document source		* Maintenance burden grew with every new document source
	\| style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" \|		\| style="vertical-align:top; padding:10px; background:#f4fbf8; color:#0f6e56;" \|
Line 651:		Line 653:

	* Layout-agnostic — works across varied formats without template changes		* Layout-agnostic — works across varied formats without template changes
	* One extraction definition handles multiple ~~vendors~~		* One extraction definition handles multiple document variants
	* Contextual understanding catches fields rule-based tools miss		* Contextual understanding catches fields rule-based tools miss
	\|-		\|-
Line 657:		Line 659:
	'''Result:''' ''A single natural language prompt replaces dozens of fragile field rules and vendor templates.''		'''Result:''' ''A single natural language prompt replaces dozens of fragile field rules and vendor templates.''
	\|}		\|}

	~~----~~