simpleRAG() is the high-level API: pass document source, chat model, and optional options — the platform handles everything else automatically. agent() is for when you need to override a specific stage that simpleRAG() does not expose.

Use agent() when you need this nested ingestion model and full retrieval configuration (retrievalAugmentor, queryRouter, and so on), not the flat options struct used by simpleRAG().

Note: Neither style replaces the other: use agent() ingestion when you want a declarative RAG setup; use documentService() when you need procedural control or reuse of intermediate arrays (documents, segments).

The agent() ingestion pipeline runs in a fixed order: Loader → Document Transformer → Splitter → Segment Transformer → Vector Store Ingestor. Each stage is optional — omit any stage to use its default behavior.

Optional ingestion.documentLoader makes the loader explicit. When omitted, the loader is inferred from ingestion.source. Three sourceType values are supported: "filesystem", "url", and "custom".

Combine ingestion.source with recursive, includePatterns, and documentLoader: { sourceType: "filesystem" }. Then ingest() runs the full pipeline.

Set ingestion.source to an HTTPS URL. Under documentLoader, use sourceType: "url" and requestOptions for connection and read timeouts.

sourceType: "custom" supplies a UDF (implementation) that returns an array of structs, each with text and metadata. Use this when content is built in CFML instead of read from disk.

Note: The document transformer receives one document struct at a time and must return a document struct (or null/empty to drop the document). It runs after loading and before splitting, so changes affect all downstream segments.

Configure ingestion.documentSplitter as a struct with chunkSize, chunkOverlap, splitterType, separators, and regexPattern. An empty struct {} uses defaults (recursive splitting, chunkSize 1000, chunkOverlap 100).

An optional ingestion.segmentTransformer UDF runs against each segment produced by the splitter. Use it to enrich segment metadata, filter out low-quality chunks, or normalise text after splitting but before embedding.

Note: The segment transformer receives one segment struct at a time (with text and metadata keys) and must return a segment struct or null to drop the segment. It runs after splitting and before the vector store ingestor.

Configure ingestion.vectorStoreIngestor with a vectorStore object and optional batchSize and continueOnError.

By default, synchronous ingestion (ingest()) keeps the request busy until indexing finishes. Asynchronous ingestion starts that work and returns a Future object right away.

Example: ingest asynchronously, then query

Retrieval is the part of RAG that selects a small amount of text from your knowledge base so the language model can ground its answer in your documents instead of relying only on parametric memory. The retrievalAugmentor struct on an agent() wires optional stages:

Typical order in config: declare queryTransformer (if any), contentAggregator (if any), contentInjector (if any), and queryRouter (with contentRetrievers). At runtime the product applies transformation, routes to one or more retrievers, aggregates chunks, then injects context for the language model.

Query transformation adjusts the text used for embedding search and context assembly. The compressing transformer rewrites the current user message using chat history so vague follow-ups (for example "Tell me more about that.") still retrieve relevant chunks.

Use queryTransformer with CHATMEMORY (for example messageWindowChatMemory) so prior turns exist to resolve references.

With contentAggregator, the same transformer composes when multiple retrievers or merged chunks are used:

If queryTransformer is omitted, retrieval uses the raw user message. The suite still expects a relevant answer when the question is explicit.

Query routing decides which retrieval path to use. queryRouter holds an array of contentRetrievers, each pointing at a vectorStore and carrying a description that characterises what that retriever covers.

A single-store, single-retriever setup is the minimal case:

You can register two or more retrievers pointing at the same vectorStore with different description, maxResults, or minScore:

Each retriever is a struct that must include a vectorStore reference for augmented retrieval.

Re-ranking and ordering results:

Together, these control precision versus recall of retrieved text.

When queryRouter lists several contentRetrievers, each may return its own set of chunks. contentAggregator (for example type: "default" with a separator) merges those strings for the generator.

The Content Injector is the component that controls how retrieved content is merged with the user's question before the LLM sees it. It sits between the retriever and the LLM, and determines the exact shape of the prompt the model receives.

contentInjector accepts two optional properties. When omitted entirely, ColdFusion uses default injection behavior (retrieved content is prepended before the user message).

Your template string can contain two Mustache placeholders:

Both placeholders are optional. If {{contents}} is absent, retrieved content is not injected. If {{userMessage}} is absent, the user's question is not included. If neither placeholder is present, ColdFusion logs a warning and falls back to default injection, the application does not crash.

The examples below use an `agent()` configured with an in-memory vector store and Mistral as the LLM. Your application will substitute its own API keys and model choices. All setup code belongs in `Application.cfc`:

When you omit contentInjector, ColdFusion automatically prepends retrieved content before the user's message. This is the simplest starting point.

Klarna is a payment service that allows you to shop online and pay for your purchases later through various payment options, including Buy Now, Pay Later (BNPL). It may perform a soft credit check upon signup, which does not affect your credit score. You can manage your Klarna purchases and outstanding balances via the Klarna app. If you need more information or have specific questions, you can visit Klarna's FAQ page or contact their Customer Service.

Use promptTemplate when you need precise control over how the LLM sees the context and the question.

To view your Adobe support cases, sign in to your Adobe account and navigate to the Support history page. There, you will find all your cases listed under the Support cases section. You can review both open and closed cases. If you want to specifically view your closed cases, select the 'Closed' option. Alternatively, you can see all cases, including both open and closed, by selecting 'All'. Please note that you cannot modify or reopen a closed case.

The metadataKeys property appends document metadata (such as the source filename) to each retrieved chunk before injection. This allows the LLM to reference or cite sources.

To resolve payment issues, check for a Billing issue alert displayed in your plan card. If it appears, your recent payment may have failed. You can resolve this by selecting 'Edit billing and payment' and then adding or updating your payment information. Make sure that your credit card details, including expiration date and billing address, are current. If you have multiple profiles associated with the same email, ensure you are managing the right account.

You can specify multiple metadata keys. ColdFusion appends each available field to the chunk text.

Common metadata keys available on ingested documents:

Key	Description
file_name	The filename of the source document
absolute_directory_path	Full directory path containing the file
absoluteFilePath	Full file path including the filename
fileSize	File size in bytes

Our product features include native support for Retrieval-Augmented Generation (RAG), which integrates information retrieval with text generation to enhance response accuracy and context. Additionally, ColdFusion 2025 includes advanced document processing capabilities, embedding, and retrieval features, and supports high-dimensional vector storage for semantic similarity search, making it ideal for AI applications.

Combine a custom template with `metadataKeys` to instruct the LLM to cite its sources.

To update your tax identification number (TIN), follow these steps based on your account type: 1. **For Individual Accounts**: Look for the section on how to update your business tax identification number, such as VAT, GST, or NIT. Instructions may be found on Adobe's support page. 2. **For Tax-Exempt Customers in North America**: Follow the step-by-step instructions provided by Adobe to place a tax-exempt order. 3. **For PayPal Users**: Ensure accurate tax information appears on future invoices by updating your VAT or tax details directly in your PayPal account settings.

Add CHATMEMORY alongside contentInjector to support follow-up questions. The injector formats context correctly on every turn; chat memory preserves conversational state.

Turn 1: The system requirements for Adobe products vary depending on the specific product you are using. Generally, you need to have the latest version of your operating system and browser for optimal functionality and security. Make sure your system meets the minimum specifications provided on Adobe's official website for the product you are interested in.Turn 2: To set up auto-renewal for your Adobe subscription, follow these steps: Sign in to your Adobe account. Navigate to the 'Billing and payment' section, where you will find an option to toggle auto-renewal on or off. If you enable auto-renewal, your subscription will automatically renew at the end of each annual term, and you will receive an annual bill until you decide to cancel or opt out.Turn 3: To convert your Creative Cloud trial to a paid subscription, follow these steps: 1. Make sure you have access to the Adobe ID used for your trial, a stable internet connection, and a valid payment method (credit card, debit card, or PayPal). 2. Go to the Creative Cloud website. 3. Sign in to your account. 4. Follow the prompts to convert your trial to a paid membership. This will allow you to continue using Adobe apps and services without interruption.

When you only need one retriever, you can use contentRetriever directly on retrievalAugmentor instead of nesting it under queryRouter. This is equivalent but more concise.

We support payment methods including credit cards (Mastercard and Visa) and PayPal. You can switch between these methods as needed.

When using multiple retrievers, a contentAggregator can merge their results with a custom separator before the injector formats the combined text.

To manage your subscriptions, you can go to your Adobe account settings. Here, you can update your credit card information which will affect all subscriptions linked to that card. Additionally, you can manage your marketing communication preferences separately from account-related messages. This gives you control over the updates and communications you receive from Adobe regarding events, app updates, news, and more.

Use multiple retrievers to search different semantic areas or apply different scoring thresholds. The injector formats the combined results as a single block.

To integrate with third-party services using Facebook, you need to enable Facebook's integration with these services first. You can do this by going to your Facebook settings, navigating to the 'Apps and Websites' section, and turning on the option for apps, websites, and games. Once this is enabled, you should be able to proceed with the integration process. Make sure to check the permissions that the third-party service requests to ensure your privacy and data security.

A queryTransformer with type "compressing" rewrites multi-turn chat history into a single self-contained query before retrieval. This improves retrieval accuracy in long conversations.

To reactivate your Adobe subscription, follow these steps: 1. **Check Payment Methods**: Ensure that the payment method on your account is valid and has sufficient funds. You can update your payment information by logging into your Adobe account. 2. **Visit Account Management**: Go to the Adobe account management page and navigate to the billing section. Here you can see if your account is inactive and follow prompts to reactivate it. 3. **Attempt Payment Again**: If your account is inactive due to failed payment attempts, you may need to attempt the payment again. This can usually be done directly from the account management section. 4. **Contact Support**: ....

This example combines every component: query transformer, multiple retrievers, aggregator, injector with metadata, and chat memory.

It seems you're having trouble signing in with your social account. Please check if your social account is connected properly. If you're receiving an error message stating 'We couldn't connect your social account,' make sure you choose the correct option when prompted. For example, if logging in with Facebook, ensure you select 'Continue as [your name]' instead of 'Not now.' If you still can't sign in, please provide more details.

The content injector works identically with all supported vector store providers. Swap provider: "INMEMORY" for provider: "qdrant" to use a persistent, production-grade store.

Point source at a directory path instead of a single file to ingest all documents at once. The injector will format content from any of those files depending on which are most relevant to the query.

It seems you're having trouble with your Adobe subscription. There could be multiple reasons for this, such as having multiple accounts, not authenticating your card details, or other issues. Please check if you are logged in with the correct Adobe ID associated with your subscription. Additionally, verify your payment details to ensure they are correctly authenticated.

No documents retrieved (high minScore): When minScore is set very high (e.g., 0.99) and no chunks meet the threshold, {{contents}} resolves to an empty string. The LLM receives the template with an empty context section and falls back to its own knowledge. The application does not throw an error.

Template missing one placeholder: A template with only {{contents}} (no {{userMessage}}) delivers context to the LLM but omits the user's actual question. A template with only {{userMessage}} delivers the question but no retrieved context, effectively bypassing RAG. Both are accepted without errors.

Template with no placeholders:

A template string with no recognized placeholders causes ColdFusion to log a warning and fall back to default injection. The application continues to run.

abc

Intelligent or multi-domain scenarios use separate VectorStore instances, each populated from different documents. description distinguishes domains (economics vs technology vs astronomy). The query-only service uses queryRouter with multiple contentRetrievers so the router can send the question to the store that matches the user intent.

The pattern has three steps:

With one retriever, routing is trivial. Every query uses that store. With multiple retrievers, the implementation uses the description values (and optional type on the router) to steer finance vs support vs internal docs questions to the appropriate index.

Note: The query-only agent() has no ingestion block. Ingestion ran separately on svcEcon and svcTech. This pattern lets you update one domain's index without rebuilding the others, and it lets different teams own different corpora independently.