LMM Labs · Supporting sources
Link and product-role verification: October 19, 2026
This register supports Surveying the AI Landscape. It follows Table 1’s ten category IDs and preserves its fifty product candidates. Each entry identifies an official source, the implementation role it supports, and a qualification relevant to selection.
These are documentation checks, not hands-on product benchmarks or rankings. The taxonomy and shortlist remain LMM Labs’ editorial synthesis. Documentation supports product capabilities, not the claim that a particular stack will succeed in a particular business. Table 2 and the worked example remain illustrative design analyses.
Dates and access. All sources below were opened and inspected on October 19, 2026. Unless a source gives its own publication date, treat it as undated living documentation retrieved on that date. Current destinations and redirects are supplied for navigation; this register does not assert that today’s page contents were unchanged at the article’s October 1 research cutoff. Public documentation is readable without a product account; using a service may require one. Features, editions, regions, pricing, licenses, and maintenance status require implementation-time confirmation.
1A Predictive model development
- scikit-learn
scikit-learn developers. (n.d.). User guide. Retrieved October 19, 2026.
Role supported: Predictive estimators, preprocessing, model selection, calibration, and evaluation.
Selection qualification: The stable documentation changes with releases; pin the library version used in a project. - XGBoost
XGBoost developers. (n.d.). XGBoost documentation. Retrieved October 19, 2026.
Role supported: Gradient-boosted tree learning for predictive modeling.
Selection qualification: Compare on the actual data and error costs; the source does not establish a universal winner. - LightGBM
LightGBM developers. (n.d.). LightGBM documentation. Retrieved October 19, 2026.
Role supported: Tree-based gradient boosting for predictive models.
Selection qualification: Efficiency claims are vendor/project descriptions, not a comparative result measured for this article. - PyTorch
PyTorch contributors. (n.d.). PyTorch documentation. Retrieved October 19, 2026.
Role supported: Tensor computation, automatic differentiation, and neural-network development.
Selection qualification: The stable entry redirected to version 2.14 during this check; choose the documentation for the installed version. - TensorFlow
TensorFlow contributors. (n.d.). TensorFlow Core guide. Retrieved October 19, 2026.
Role supported: Neural-model construction, training, and deployment tooling.
Selection qualification: A neural model is an option, not a requirement for every predictive task.
1B Generative model APIs
- OpenAI API
OpenAI. (n.d.). OpenAI API documentation. Retrieved October 19, 2026.
Role supported: API access to generative models and application tools.
Selection qualification: Model capabilities, prices, retention terms, and regional availability must be checked for the selected service. - Claude API
Anthropic. (n.d.). Intro to Claude. Retrieved October 19, 2026.
Role supported: Direct model prompting through the Claude API.
Selection qualification: The direct Messages API is distinct from the Agent SDK and hosted Managed Agents. - Gemini API
Google. (n.d.). Gemini API. Retrieved October 19, 2026.
Role supported: Generative-model APIs for text and supported multimodal tasks.
Selection qualification: Select an appropriate model and API; modality and interface support vary. - Mistral API
Mistral AI. (n.d.). Mistral AI documentation. Retrieved October 19, 2026.
Role supported: Model inference APIs and developer tooling.
Selection qualification: Use the inference/API documentation, rather than assuming the consumer interface is the application API. - Cohere API
Cohere. (n.d.). An overview of the Cohere platform. Retrieved October 19, 2026.
Role supported: Text generation through the Chat endpoint, alongside retrieval-related models.
Selection qualification: Generation, embeddings, and reranking are different capabilities; choose the needed endpoints.
2A Search and evidence storage
- Azure AI Search
Microsoft. (n.d.). What is Azure AI Search?. Retrieved October 19, 2026.
Role supported: Managed full-text, vector, and hybrid retrieval.
Selection qualification: Permissions and applicable revisions must be implemented and tested; retrieval similarity alone is insufficient. - Amazon OpenSearch Service
Amazon Web Services. (n.d.). What is Amazon OpenSearch Service?. Retrieved October 19, 2026.
Role supported: Managed OpenSearch deployment, search, and access-control capabilities.
Selection qualification: Provisioned domains and serverless configurations have different operating and integration requirements. - Elasticsearch
Elastic. (n.d.). Search with Elasticsearch. Retrieved October 19, 2026.
Role supported: Full-text, vector, semantic, and hybrid search approaches.
Selection qualification: Deployment and subscription choices affect available controls and operating responsibility. - Pinecone
Pinecone. (n.d.). Pinecone Database. Retrieved October 19, 2026.
Role supported: Vector/semantic retrieval, metadata filtering, and related search capabilities.
Selection qualification: Metadata filters and namespaces need an application-level authorization design. - Weaviate
Weaviate. (n.d.). Weaviate Database. Retrieved October 19, 2026.
Role supported: Vector storage and semantic/hybrid search for retrieval applications.
Selection qualification: Distinguish the database, managed cloud service, and additional agent services.
2B RAG and knowledge apps
- LlamaIndex
LlamaIndex. (n.d.). LlamaIndex framework documentation. Retrieved October 19, 2026.
Role supported: Document ingestion, indexing, retrieval, and RAG application components.
Selection qualification: A framework still needs model services, storage, access controls, and deployment choices. - Haystack
deepset. (n.d.). Introduction to Haystack. Retrieved October 19, 2026.
Role supported: Composable pipelines, document stores, retrieval, and generation integrations.
Selection qualification: Pipeline and agent modes differ in execution control; use the mode needed by the application. - LangChain
LangChain. (n.d.). LangChain overview. Retrieved October 19, 2026.
Role supported: Model, tool, and embedding integrations for AI applications.
Selection qualification: It is broader than RAG; its placement identifies one implementation role, not an exclusive product category. - Dify
Dify. (n.d.). Dify documentation. Retrieved October 19, 2026.
Role supported: Visual AI application/workflow assembly and knowledge integration.
Selection qualification: Cloud and self-hosted editions have different administration and operational requirements. - Microsoft GraphRAG
Microsoft. (n.d.). GraphRAG documentation. Retrieved October 19, 2026.
Role supported: Knowledge-graph extraction, community summaries, and graph-based retrieval/synthesis.
Selection qualification: Extra indexing work is justified only when it improves the required questions; it is not a default replacement for conventional RAG.
3A Agent construction and tool control
- OpenAI Agents SDK
OpenAI. (n.d.). OpenAI Agents SDK. Retrieved October 19, 2026.
Role supported: Agent loops, function tools, guardrails, sessions, and tracing.
Selection qualification: The SDK and a direct model API are different implementation choices; test approval and resume behavior. - Claude Agent SDK
Anthropic. (n.d.). Agent SDK overview. Retrieved October 19, 2026.
Role supported: Programmable Claude Code agent loop with tools, permissions, hooks, and sessions.
Selection qualification: The SDK runs a Claude Code process; do not confuse it with a thin API client or hosted Managed Agents. - Strands Agents
Strands Agents. (n.d.). Get started. Retrieved October 19, 2026.
Role supported: Model-driven agent construction with tools, session management, and runtime controls.
Selection qualification: Python and TypeScript feature availability can differ; Bedrock is a provider option rather than the whole SDK. - Google ADK
Google. (n.d.). Agent Development Kit documentation. Retrieved October 19, 2026.
Role supported: Agent development with tools, workflows, state, and runtime integration.
Selection qualification: The development kit is separate from the managed Google Cloud deployment platform. - Microsoft Agent Framework
Microsoft. (n.d.). Microsoft Agent Framework overview. Retrieved October 19, 2026.
Role supported: Agents, tool integrations, sessions, and explicit workflow execution paths.
Selection qualification: Language and release maturity differ; the overview identifies preview limitations for some implementations.
3B Business integration and automation
- n8n
n8n. (n.d.). Build your first workflow. Retrieved October 19, 2026.
Role supported: Trigger-driven workflows connecting API steps and stored data.
Selection qualification: Choose cloud or self-hosting, then test credentials, delivery, and retries for the required connectors. - Power Automate
Microsoft. (n.d.). Explore the Power Automate home page. Retrieved October 19, 2026.
Role supported: Business workflows across applications and services, including approval-related features.
Selection qualification: Connector availability, licensing, and authentication scopes must be checked for the actual flow. - Make
Make. (n.d.). Help Center: Get started. Retrieved October 19, 2026.
Role supported: Visual automation scenarios, integrations, and error-handling guidance.
Selection qualification: Verify exact connector operations and replay/retry behavior before allowing writes. - Zapier
Zapier. (n.d.). Get started with Zapier. Retrieved October 19, 2026.
Role supported: Configured app-to-app workflows using triggers and actions.
Selection qualification: The integration catalog does not guarantee that every required operation is available in a chosen plan. - Pipedream Connect
Pipedream. (n.d.). Connect overview. Retrieved October 19, 2026.
Role supported: Embedded integrations, managed authentication, API proxying, and tools.
Selection qualification: Connect is not the retiring Workflows product; see the separate lifecycle notice below.
4A Stateful execution and recovery
- LangGraph
LangChain. (n.d.). LangGraph overview. Retrieved October 19, 2026.
Role supported: Stateful orchestration combining deterministic code and model-driven steps.
Selection qualification: Persistence and human review still require correct state/storage and application configuration. - Temporal
Temporal Technologies. (n.d.). Temporal Workflow Execution overview. Retrieved October 19, 2026.
Role supported: Durable workflow state, replay, and recovery across failures and waits.
Selection qualification: Durable orchestration does not by itself make an external business write safe to repeat. - AWS Step Functions
Amazon Web Services. (n.d.). What is Step Functions?. Retrieved October 19, 2026.
Role supported: State-machine workflows, service integration, and execution inspection.
Selection qualification: Standard and Express workflows differ in duration and delivery semantics; select the appropriate type. - Azure Durable Functions
Microsoft. (n.d.). Durable Functions overview. Retrieved October 19, 2026.
Role supported: Stateful serverless orchestration with checkpoints, retries, and recovery.
Selection qualification: Orchestrator, activity, and entity code have distinct responsibilities; test external side effects. - Prefect
Prefect. (n.d.). Introduction. Retrieved October 19, 2026.
Role supported: Python-oriented workflow orchestration and deployable flows.
Selection qualification: Scheduling a pipeline does not validate its model or guarantee correct business outcomes.
4B Managed AI application platforms
- Amazon Bedrock
Amazon Web Services. (n.d.). Amazon Bedrock overview. Retrieved October 19, 2026.
Role supported: Managed access to foundation models and supporting application services.
Selection qualification: Confirm service-specific region, model, identity, storage, and operational requirements. - Microsoft Foundry
Microsoft. (n.d.). What is Microsoft Foundry?. Retrieved October 19, 2026.
Role supported: Managed AI development and application services.
Selection qualification: Some capabilities are in preview; use feature-specific availability and deployment documentation. - Gemini Enterprise Agent Platform
Google Cloud. (n.d.). Agent Platform overview. Retrieved October 19, 2026.
Role supported: Managed agent/model development, runtime, governance, and evaluation services.
Selection qualification: The current overview was updated October 16, 2026; it is current navigation, not an archive of the article cutoff. - IBM watsonx.ai
IBM. (n.d.). IBM watsonx.ai. Retrieved October 19, 2026.
Role supported: AI model development, inference, RAG, and lifecycle capabilities.
Selection qualification: This is an official product overview, not independent comparative testing; watsonx Orchestrate is a separate integration choice. - Databricks AI platform
Databricks. (n.d.). Machine learning on Databricks; Develop agents on Databricks. Retrieved October 19, 2026.
Role supported: Integrated model development, serving, and monitoring, with related agent services.
Selection qualification: The linked documentation is the AWS edition; use the appropriate cloud and feature documentation for implementation. Additional official source: Develop agents on Databricks (updated October 13, 2026).
5A Tracing and evaluation platforms
- LangSmith
LangChain. (n.d.). LangSmith observability. Retrieved October 19, 2026.
Role supported: Tracing, feedback, monitoring, and evaluation workflows.
Selection qualification: A complete trace requires the application to instrument the relevant retrieval and tool steps. - Langfuse
Langfuse. (n.d.). Langfuse overview. Retrieved October 19, 2026.
Role supported: Tracing, prompt management, evaluation, and dataset workflows.
Selection qualification: Cloud and self-hosted operation differ; review sensitive-data handling and export requirements. - Arize Phoenix
Arize AI. (n.d.). What is Arize Phoenix?. Retrieved October 19, 2026.
Role supported: Tracing, evaluation, datasets, experiments, and feedback.
Selection qualification: Phoenix and the managed enterprise Arize AX platform are distinct offerings. - Braintrust
Braintrust. (n.d.). Get started with Braintrust. Retrieved October 19, 2026.
Role supported: Application traces, evaluation experiments, datasets, and human annotations.
Selection qualification: Evaluate retention, deployment, access controls, and export against the application requirements. - W&B Weave
Weights & Biases / CoreWeave. (n.d.). W&B Weave. Retrieved October 19, 2026.
Role supported: Observability, tracing, scoring, and evaluation of model applications.
Selection qualification: The former Weave documentation address redirects to CoreWeave documentation.
5B Executable evaluation and failure tests
- Promptfoo
Promptfoo. (n.d.). Intro. Retrieved October 19, 2026.
Role supported: Executable evaluations, output assertions, and red-team tests.
Selection qualification: Passing configured checks is not proof that unspecified failure modes are covered. - DeepEval
Confident AI. (n.d.). DeepEval 5-min Quickstart. Retrieved October 19, 2026.
Role supported: Local evaluation tests, agent trajectories, and component-level diagnosis.
Selection qualification: Choose metrics for the failure being tested and validate model-judge judgments against expert review. - Ragas
Ragas developers. (n.d.). Ragas documentation. Retrieved October 19, 2026.
Role supported: RAG evaluation metrics and related testing workflows.
Selection qualification: The current library is broader than the original RAGAs paper; metric definitions depend on the selected version. - Giskard
Giskard AI. (n.d.). Giskard documentation: testing framework. Retrieved October 19, 2026.
Role supported: ML/LLM scanning, test suites, and RAG evaluation tooling.
Selection qualification: The documentation now explicitly resides at a legacy address; confirm maintenance and package compatibility before adopting it. - TruLens
TruLens contributors. (n.d.). Getting started. Retrieved October 19, 2026.
Role supported: Feedback metrics, retrieval evaluation, tracing, and test integrations.
Selection qualification: Its evaluation and tracing features overlap categories 5A and 5B; placement is not exclusive.
Supporting implementation and lifecycle sources
- Amazon Web Services. (n.d.). Retrieve data and generate AI responses with Amazon Bedrock Knowledge Bases. Retrieved October 19, 2026. Supports the retrieval and cited-generation option in Table 2 and the worked case. A customer-managed knowledge base needs a supported vector-store configuration; managed and customer-managed offerings differ.
- Microsoft. (2026, August 31). Hybrid search using vectors and full-text search in Azure AI Search. Retrieved October 19, 2026. Supports combining keyword and vector retrieval; it does not establish the worked case’s comparative answer quality.
- OpenAI. (n.d.). Structured model outputs. Retrieved October 19, 2026. Supports schema-constrained output while explicitly warning that outputs can still contain mistakes.
- Pipedream. (n.d.). Workflows and String are shutting down on March 31, 2027. Retrieved October 19, 2026. The notice explicitly states that Connect is unaffected.
How to use this register. Follow the article’s flowchart: identify the required role, use its Table 1 category to find sources here, then test the actual products against the same acceptance criteria. Confirm connector operations, permission enforcement, recovery, model-judge reliability, cost, and maintenance obligations in a pilot. A documentation link is evidence of a described capability, not proof of fitness or production readiness.
Prepared by LMM Labs in collaboration with ChatGPT. © 2026 LMM Labs. Product names and third-party documentation remain the property of their respective owners.