Glossary · 43 terms
Plain-English definitions for the language we use.
A working glossary across the eCirql practice. AI engineering terms (inference, retrieval-augmented generation, function calling, agentic), integration and iPaaS terms (process flow, connector, idempotency, dead-letter queue), and commerce terms (ERP, WMS, OMS, marketplace of record, OneWorld, SuiteScript). Definitions are written to be useful in buyer conversations, not to win Scrabble points.
AI engineering
AI engineering · 17 terms
- AIiPaaS AI Integration Partner as a Service
-
AI Integration Partner as a Service. A category of AI-native integration tooling that combines a traditional iPaaS with large language model inference, retrieval-augmented generation, function calling and agentic planning. Coined by eCirql in 2026; the first product in the category is PatchBuddy.
- Inference
-
The act of running a trained model against an input to produce an output. In production AI, every model call is an inference and carries cost, latency and accuracy implications that have to be reasoned about per turn rather than assumed.
- Large language model LLM
-
A foundation model trained on a large text corpus and capable of generating text, reasoning over context, calling tools and following instructions. Frontier LLMs from providers like Anthropic, OpenAI and Google sit at one end; smaller open-weight LLMs sit at the other; production AI typically uses both, selected per task.
- Foundation model
-
A large pre-trained model intended as the base for many downstream tasks rather than for one purpose. Frontier LLMs are foundation models; image models, speech models and multimodal models all qualify.
- Frontier model
-
An informal term for the most capable foundation models available at any given time, typically from providers operating at the leading edge of training compute. The frontier shifts; the engineering discipline of choosing the right model for the right task does not.
- Retrieval-augmented generation RAG
-
A pattern that grounds an LLM in external data by retrieving relevant passages at inference time and feeding them into the prompt. Avoids the hallucination failure mode of asking the model to recall facts from its training corpus, and lets the answer reflect data that postdates the model's training cutoff.
- Vector embedding embedding
-
A numerical representation of a piece of content (a paragraph, an image, a code snippet) in a high-dimensional space where semantically similar content sits close together. Generated by a dedicated embedding model and stored in a vector database for similarity search.
- Vector search
-
Similarity search over vector embeddings, returning the items closest to a query vector. The retrieval half of RAG. Production setups pair vector search with a reranker model to improve precision before the top-k candidates reach the prompt.
- Function calling tool use
-
A pattern where the LLM emits a structured call to a named function with typed arguments, and the calling environment validates, executes and returns the result. The backbone of agentic systems that take real actions; without it, the model can only describe what it would do.
- Structured output
-
Constraining a model's output to a predefined schema (JSON, a typed function call, an enumerated decision). Lets downstream code handle the response without string parsing and lets the calling environment reject and repair non-conforming outputs before they reach a system.
- Agentic
-
Describes AI systems that plan, take actions, observe results and adapt, rather than producing a single response and stopping. An agentic system decomposes a goal into steps, executes those steps against tools and surfaces decisions to a human at the points that need it.
- Context window
-
The maximum number of tokens an LLM can consider in a single inference call. Larger context windows allow more retrieved material, longer conversation history and more code to be in scope, at the cost of higher latency and price per call.
- Token
-
The unit of text an LLM reads and writes. Roughly four characters of English text per token in most tokenisers. Inference is priced and budgeted in tokens, so prompt engineering and retrieval design are partly about staying inside a token budget.
- Fine-tuning
-
Updating the weights of a pre-trained model on a smaller, task-specific dataset so it performs better on that task. Often a substitute for very large prompts; sometimes a way to teach the model a private taxonomy or tone of voice it would not otherwise know.
- Guardrails
-
Engineered constraints that sit between a model and a real-world action: input filtering, output validation, PII redaction, content classification, rate limits, audit logging. The bridge between a model that can do anything and a system that should only do specific things.
- Multimodal
-
Operating across more than one input or output modality: text, image, audio, video. Multimodal LLMs let an agent reason over a screenshot or a PDF the same way it reasons over text.
- Evaluation harness evals
-
A test suite for an AI system: a corpus of representative inputs, expected outputs (or scoring rubrics), and automated runs that score the system. The AI-engineering equivalent of regression tests; without it, model upgrades and prompt changes ship blind.
Integration & iPaaS
Integration & iPaaS · 11 terms
- iPaaS Integration Platform as a Service
-
A platform for building and operating integrations between business systems, usually with a visual canvas, a connector library, a scheduler, monitoring and audit. Patchworks is the iPaaS we deliver on most often.
- Process flow
-
The unit of work in Patchworks: a directed graph of steps that moves data between systems. One process flow handles one logical job (orders Shopify-to-NetSuite, inventory NetSuite-to-Shopify). An integration estate is a portfolio of process flows, not a single canvas.
- Connector
-
A reusable piece of integration plumbing that knows how to talk to one specific external system (Shopify, NetSuite, Amazon, etc). Provides typed inputs and outputs to process flows above it. iPaaS platforms ship a marketplace of pre-built connectors; missing ones can be authored bespoke.
- Webhook
-
An HTTP callback fired by a source system when an event occurs. Push-based, low-latency, and idempotency-sensitive. Where a source supports webhooks, integrations should use them in preference to polling.
- Polling
-
Periodically asking a source system for changes since a known watermark. Higher latency than webhooks and heavier on the source's rate limits, but the only option where webhooks are not available or not trustworthy.
- Idempotency
-
The property that running an operation more than once has the same effect as running it once. Critical for integrations because retries, replays and double-deliveries are routine; a flow that creates a duplicate order on a retried webhook is broken even if it works on the happy path.
- Dead-letter queue DLQ
-
A holding area for events or messages that could not be processed successfully and have exhausted their retry budget. Lets operations review and resolve failures without blocking the rest of the queue, and turns an outage into a triage exercise rather than a silent data-loss event.
- EDI Electronic Data Interchange
-
A family of standards (EDIFACT, X12 and others) for exchanging business documents in fixed structured formats. Common in retail, logistics and manufacturing where two trading partners need a stable wire format that survives organisational changes on both sides.
- SFTP Secure File Transfer Protocol
-
File transfer over SSH. Often the integration surface for legacy systems that cannot expose a modern API but can drop a CSV or fixed-width file on a schedule. Modern iPaaS platforms treat SFTP as a first-class source.
- Patchworks Tapestry
-
The previous-generation Patchworks integration platform. Replaced by Patchworks Core; deprecation has been signalled by Patchworks and migration to Core is the active path for existing Tapestry estates.
- Patchworks Core
-
Patchworks' current-generation integration platform. The destination for estates migrating off Tapestry, and the platform where new integrations on the Patchworks stack are built today.
Commerce, ERP & marketplace
Commerce, ERP & marketplace · 15 terms
- ERP Enterprise Resource Planning
-
The system of record for a business's financial, operational and stock data. NetSuite, Microsoft Dynamics Business Central and Sage 200 are common ERPs in the merchants we work with. Integrations typically treat the ERP as the source of truth for inventory, customers and posted finance.
- WMS Warehouse Management System
-
The system that runs a physical warehouse: receive, putaway, pick, pack, dispatch, count. Common WMS platforms in our estate include Descartes Peoplevox, Veeqo and ShipBob. WMS-to-ERP and WMS-to-storefront integrations are some of the most common flow shapes we ship.
- 3PL Third-Party Logistics
-
A logistics provider that runs fulfilment on behalf of the merchant. May operate its own WMS or use a customer-supplied one. Integrations to a 3PL look like integrations to a WMS, with stricter SLA awareness and tighter data exchange contracts.
- OMS Order Management System
-
A layer that orchestrates orders across multiple channels and fulfilment locations. Decides where to source stock from, when to split a shipment, when to backorder. Sits between the storefront and the ERP / WMS in operations that need this routing as a first-class concern.
- PIM Product Information Management
-
The system of record for product master data, attributes, locales and channel-specific enrichment. Akeneo, Salsify and Plytix are common PIM platforms. The integration job is usually to syndicate enriched data out to storefronts and marketplaces.
- DXP Digital Experience Platform
-
A broader category covering CMS, personalisation, search and commerce in one platform stack. Less precise than ERP or WMS; in our work, DXP-aligned platforms get treated like a CMS plus storefront combined.
- Settlement
-
The financial reconciliation report a marketplace sends after a payout period, listing every order, refund, fee, adjustment and chargeback it has applied. Reconciling settlement against the underlying ERP or accounting record is the difference between marketplace selling that is auditable and marketplace selling that is a guess.
- Marketplace
-
A third-party sales channel where the merchant lists products and the marketplace owns the customer relationship and (often) the payment. Amazon, Mirakl operators, eBay, TikTok Shop. Marketplace integrations need explicit handling of category attributes, SLA-bound dispatch and settlement reconciliation.
- FBA Fulfilled by Amazon
-
Amazon's fulfilment model in which the merchant ships stock to Amazon's warehouses and Amazon handles pick, pack and dispatch. FBA inventory needs to be modelled as a distinct location in the ERP, with its own availability rules and movement events.
- Marketplace of record MoR
-
A regulatory model in which the marketplace collects and remits VAT or sales tax on the seller's behalf for orders shipped to certain jurisdictions. Changes the shape of the seller's tax obligation and how marketplace orders should post to the ERP.
- NetSuite OneWorld
-
NetSuite's multi-subsidiary edition. Required when a group trades through multiple legal entities with separate ledgers and consolidated reporting. Integrations need to be subsidiary-aware: an order from a UK storefront has to post to the UK subsidiary, not the parent.
- SuiteScript
-
NetSuite's JavaScript-based customisation language. Used for bespoke logic on records, scheduled tasks, RESTlets that expose custom HTTP endpoints, and integration callbacks. Most non-trivial NetSuite integrations involve at least some SuiteScript.
- SuiteQL
-
NetSuite's SQL dialect for querying records. The escape hatch when SuiteScript record APIs are too slow or too rigid for a reporting use case, and the language finance teams actually run their dashboards in.
- B2B commerce
-
Business-to-business ecommerce: customer accounts, tiered pricing, credit balances, invoice settlement, purchase-order workflows. SparkLayer B2B, Shopify B2B and BigCommerce B2B are the platforms we see most often. B2B integration is materially different from D2C and needs its own scoping.
- D2C commerce DTC
-
Direct-to-consumer ecommerce: a brand selling its own products to end customers, usually through its own storefront and marketing channels. The simpler counterpart to B2B but no less rigorous on integration discipline.
Get in touch
Tell us what you’re trying to connect.
And what’s in the way. We will tell you whether we are the right people to do it. Drop us a line below, or open the chat in the corner of the screen.
Direct: [email protected]
Message sent
Thanks. We’ll be in touch.
Your message is in our inbox. A human will read it and reply within a working day. For anything urgent, drop us a line at [email protected] or open the chat in the corner of the screen.