Skip to content

Kit · Tools

Tools and radar

Every product on the map, and the libraries I reach for around it. The rings are my own experience in real systems, not an industry ranking.

Radar

Radar of the tools I use10 rated tools in four rings, Adopt in the centre to Hold at the edge. The numbered list beside it reads the same.AdoptTrialAssessHoldAI and dataPlatformSecurity and identityIntegration and operations1. Instructor: Adopt, AI & Models12. LangGraph: Adopt, AI & Models23. pgvector: Adopt, Data & Databases34. OpenTelemetry: Adopt, Observability & Ops45. Ragas: Adopt, Observability & Ops56. LlamaIndex: Trial, Data & Databases67. OpenAI Agents SDK: Trial, AI & Models78. Unseen UI: Trial, AI & Models89. LiteLLM: Trial, Networking910. Guardrails: Assess, Security10
Fig. 1 · 10 rated tools. Adopt in the centre, Hold at the edge.

Adopt 5

I use it in production and would again.

  1. 1 Instructor
  2. 2 LangGraph
  3. 3 pgvector
  4. 4 OpenTelemetry
  5. 5 Ragas

Trial 4

In a real project, not yet proven at scale.

  1. 6 LlamaIndex
  2. 7 OpenAI Agents SDK
  3. 8 Unseen UI
  4. 9 LiteLLM

Assess 1

Read, prototyped, not yet trusted with a workload.

  1. 10 Guardrails

Hold 0

Tried it; would not start a new project with it.

Nothing in this ring yet.

Quadrants: AI and data · Platform · Security and identity · Integration and operations

The list

Open a row for where it fits, where it does not and what to use instead. Map links open the component with that stack selected.

  • 1InstructorStructured outputs for LLMs using Pydantic models.Adopt AI & ModelsAny stack
    Fits
    Typed outputs with retries in a few lines.
    Does not fit
    Streaming-heavy UIs where partial objects matter.
    Alternatives
    Native structured outputs
  • 2LangGraphBuild stateful, multi-agent applications as explicit graphs.Adopt AI & ModelsOpen source
    Fits
    Explicit graphs when routing logic is the hard part.
    Does not fit
    Simple single-agent apps; the ceremony is not worth it.
    Enterprise concerns
    Checkpointer backends and state size need an operational owner.
    Alternatives
    OpenAI Agents SDK, Semantic Kernel

    On the map: Agent

  • 3pgvectorOpen-source vector similarity search for PostgreSQL.Adopt Data & DatabasesOpen source
    Fits
    Keeping embeddings next to relational data for most workloads.
    Does not fit
    Hundreds of millions of vectors with strict latency targets.
    Enterprise concerns
    Index tuning (HNSW parameters) needs a DBA conversation.
    Alternatives
    Qdrant, Azure AI Search

    On the map: Vector Index

  • 4OpenTelemetryObservability for modern systems, with GenAI semantic conventions.Adopt Observability & OpsOpen source
    Fits
    One tracing standard across agents, tools and HTTP.
    Does not fit
    Nothing; adopt it even for a prototype.
    Enterprise concerns
    GenAI semantic conventions are still stabilising; expect attribute renames.
    Alternatives
    Vendor SDKs

    On the map: Observability

  • 5RagasEvaluation framework for retrieval-augmented generation pipelines.Adopt Observability & OpsAny stack
    Fits
    Retrieval and faithfulness metrics in CI against a golden set.
    Does not fit
    Metrics without a golden set; it will not invent ground truth.
    Enterprise concerns
    LLM-as-judge cost on every CI run; sample.
    Alternatives
    DeepEval, Custom evals
  • 6LlamaIndexData framework for connecting LLMs to your documents and data.Trial Data & DatabasesOpen source
    Fits
    Fast ingestion and retriever experiments over many document types.
    Does not fit
    Teams that already own a retrieval pipeline; it adds a second abstraction.
    Enterprise concerns
    Version drift between integrations; pin aggressively.
    Alternatives
    LangChain, Hand-rolled with pgvector

    On the map: Retrieval

  • 7OpenAI Agents SDKLightweight framework for multi-agent workflows with handoffs and guardrails.Trial AI & ModelsAny stack
    Fits
    Smallest API for supervisor and hand-off patterns with built-in tracing.
    Does not fit
    Provider-neutral platforms; it leans OpenAI.
    Enterprise concerns
    Vendor coupling in orchestration code.
    Alternatives
    LangGraph
  • 8Unseen UIAccessible React components for AI interfaces, by StackUnseen.Trial AI & ModelsAny stack
    Fits
    Chat, tool calls, approvals and streaming as controlled React components.
    Does not fit
    Non-React stacks.
    Enterprise concerns
    Alpha releases; the API still moves between versions.
    Alternatives
    assistant-ui, Vercel AI Elements
  • 9LiteLLMUnified LLM API interface across providers.Trial NetworkingOpen source
    Fits
    Provider abstraction plus a proxy with cost tracking and rate limits.
    Does not fit
    Single-provider teams; it is one more hop.
    Enterprise concerns
    The proxy is in your data path; treat it as production infrastructure.
    Alternatives
    Portkey, Direct SDKs

    On the map: LLM Gateway

  • 10GuardrailsValidate and control LLM outputs with reusable validators.Assess SecurityAny stack
    Fits
    Turning output policies into testable validators.
    Does not fit
    Latency-sensitive paths; validators add round trips.
    Enterprise concerns
    Validator hub supply chain; review what you install.
    Alternatives
    Instructor + Pydantic, Custom validators
On the map, not rated yet43 products I have not put a ring on
  • ·AI Agent FrameworkMosaic AI Agent Framework: agents authored in code, evaluated and served next to the data, traced with MLflow.Not rated AI & ModelsDatabricks

    On the map: Agent

  • ·AI Search indexersPull from Blob, SQL and SharePoint with built-in skills for chunking and vectorising.Not rated Data & DatabasesAzure

    On the map: Ingestion

  • ·Airflow · UnstructuredScheduled pipelines for parsing, chunking and embedding; full control of chunking.Not rated Data & DatabasesOpen source

    On the map: Ingestion

  • ·Azure OpenAIOpenAI models in your tenant and region, with private endpoints.Not rated AI & ModelsAzure

    On the map: Model

  • ·Cosmos DBPer-user partitions for threads and memory, globally distributed.Not rated Data & DatabasesAzure

    On the map: Memory

  • ·Declarative PipelinesLakeflow Connect pulls from SaaS sources such as SharePoint; Lakeflow Declarative Pipelines (formerly DLT) chunk and embed; Change Data Feed keeps indexes current.Not rated Data & DatabasesDatabricks

    On the map: Ingestion

  • ·Delta · Unity CatalogGoverned tables and volumes; permissions travel with the data.Not rated Data & DatabasesDatabricks

    On the map: Sources

  • ·Foundry Agent ServiceAzure AI Foundry Agent Service: a managed runtime with threads, tools and Entra-scoped connections.Not rated AI & ModelsAzure

    On the map: Agent

  • ·LakebaseManaged Postgres next to the lakehouse for agent state and history.Not rated Data & DatabasesDatabricks

    On the map: Memory

  • ·Logic Apps · TeamsApproval cards in Teams that resume the workflow when answered.Not rated AI & ModelsAzure

    On the map: Human review

  • ·Model ServingFoundation Model APIs plus external models, governed through the AI Gateway.Not rated AI & ModelsDatabricks

    On the map: Model

  • ·Postgres · S3 · wikiWhatever already holds the knowledge; connectors are yours to write.Not rated Data & DatabasesOpen source

    On the map: Sources

  • ·Redis · PostgresRedis for working state, Postgres for durable history and facts.Not rated Data & DatabasesOpen source

    On the map: Memory

  • ·Semantic rankerHybrid queries with Microsoft's reranker and security trimming in one call.Not rated AI & ModelsAzure

    On the map: Retrieval

  • ·SharePoint · BlobMost enterprise documents already live in Microsoft 365, permissions included.Not rated Data & DatabasesAzure

    On the map: Sources

  • ·Slack · Jira approvalApprovals land where reviewers already work; the agent waits on a callback.Not rated AI & ModelsOpen source

    On the map: Human review

  • ·Vector Search hybridHybrid keyword and vector queries with reranking. Per-document access comes from filter columns populated at ingestion, applied for the calling identity.Not rated AI & ModelsDatabricks

    On the map: Retrieval

  • ·vLLM · open weightsYour GPUs, your data residency; capacity planning becomes your problem.Not rated AI & ModelsOpen source

    On the map: Model

  • ·API ManagementPolicies for JWT validation, quotas and versioning; it can also act as an AI gateway.Not rated NetworkingAzure

    On the map: API Gateway

  • ·APIM AI gatewayToken-based limits, semantic caching and load balancing across Azure OpenAI deployments.Not rated NetworkingAzure

    On the map: LLM Gateway

  • ·Container AppsScale-to-zero containers with built-in auth integration to Entra ID.Not rated Compute & InfraAzure

    On the map: Web App

  • ·Databricks AppsApps hosted next to the data, with workspace identity and Unity Catalog access built in.Not rated Compute & InfraDatabricks

    On the map: Web App

  • ·Front Door · WAFGlobal edge, managed WAF rule sets and bot protection in one service.Not rated NetworkingAzure

    On the map: Edge

  • ·Kong · EnvoyPlugins for auth, rate limits and request logging; runs anywhere.Not rated NetworkingOpen source

    On the map: API Gateway

  • ·Mosaic AI GatewayRate limits, usage tracking, guardrails and payload logging on every serving endpoint.Not rated NetworkingDatabricks

    On the map: LLM Gateway

  • ·Next.js on K8sContainerised frontend; streaming responses need proxies that do not buffer SSE.Not rated Compute & InfraOpen source

    On the map: Web App

  • ·Nginx · ModSecurityReverse proxy with OWASP CRS rules. You own the rule tuning and the DDoS response.Not rated NetworkingOpen source

    On the map: Edge

  • ·Entra ID · OBOFirst-class on-behalf-of flow, so downstream Graph and API calls carry the user.Not rated Identity & AccessAzure

    On the map: Identity

  • ·HashiCorp VaultDynamic, short-lived credentials; the strongest model, and a service you must run.Not rated SecurityOpen source

    On the map: Secrets

  • ·Key VaultSecrets and keys read by managed identity, so config holds no credentials.Not rated SecurityAzure

    On the map: Secrets

  • ·KeycloakOIDC provider with token exchange; on-behalf-of flows are possible but you configure them.Not rated Identity & AccessOpen source

    On the map: Identity

  • ·OpenSearch · LokiAppend-only log store; immutability and retention are your configuration.Not rated SecurityOpen source

    On the map: Audit

  • ·Purview auditPurview audit and Log Analytics hold the record; Sentinel is the SIEM that correlates it with sign-ins.Not rated SecurityAzure

    On the map: Audit

  • ·Secret scopesWorkspace secrets, optionally backed by Azure Key Vault.Not rated SecurityDatabricks

    On the map: Secrets

  • ·System tablesAudit, lineage and access logs queryable as tables in Unity Catalog.Not rated SecurityDatabricks

    On the map: Audit

  • ·Workspace SSO · OAuthUsers and service principals from your IdP; permissions resolve in Unity Catalog.Not rated Identity & AccessDatabricks

    On the map: Identity

  • ·App InsightsOpenTelemetry-native; AI Foundry adds tracing and evaluation views.Not rated Observability & OpsAzure

    On the map: Observability

  • ·MCP servers on K8sOne server per system of record, deployed and scaled like any service.Not rated IntegrationOpen source

    On the map: Tools

  • ·MCP via APIMExpose existing APIs as MCP tools behind API Management policies.Not rated IntegrationAzure

    On the map: Tools

  • ·MLflow TracingTraces, evaluation runs and judges in the same place as the model versions.Not rated Observability & OpsDatabricks

    On the map: Observability

  • ·UC functions · MCPUnity Catalog functions and managed MCP servers as governed tools.Not rated IntegrationDatabricks

    On the map: Tools