Package com.composum.ai.backend.slingbase.impl

Class RAGServiceImpl

java.lang.Object

com.composum.ai.backend.slingbase.impl.RAGServiceImpl

All Implemented Interfaces:

RAGService

public class RAGServiceImpl
extends Object
implements RAGService

Basic services for retrieval augmented generation (RAG).

Field Summary

Fields

Modifier and Type	Field	Description
protected AIConfigurationService	aiConfigurationService
protected GPTChatCompletionService	chatCompletionService
protected GPTEmbeddingService	embeddingService
protected ApproximateMarkdownService	markdownService
protected PageCachedValueService	pageCachedValueService
protected AtomicLong	requestCounter

Constructor Summary

Constructors

Constructor	Description
RAGServiceImpl()

Method Summary

Modifier and Type	Method	Description
List<String>	collectSearchKeywords(String querytext, org.apache.sling.api.resource.Resource rootResource)	Processes a query to have the AI suggest a couple of search keywords for use with the other methods that might find the most relevant results.
protected @NotNull List<String>	containsQuery(@NotNull org.apache.sling.api.resource.Resource root, @NotNull String querytext, int restOfLimit)
protected GPTEmbeddingService.EmbeddingsCache	getEmbeddingsCache(Map<String,org.apache.sling.api.resource.Resource> textToResource)
protected String	normalize(String querytext)	Turn it into a query for the words mentioned in there - that is, remove all meta characters for CONTAINS queries: AND, OR, words prefixed with -, quotes, backslashes.
List<org.apache.sling.api.resource.Resource>	orderByEmbedding(String querytext, List<org.apache.sling.api.resource.Resource> resources, @NotNull org.apache.sling.api.SlingHttpServletRequest request, @NotNull org.apache.sling.api.SlingHttpServletResponse response, @NotNull org.apache.sling.api.resource.Resource rootResource)	Finds the resources whose markdown approximation has embeddings that are the most similar to the querytext embedding.
String	ragAnswer(String querytext, List<org.apache.sling.api.resource.Resource> resources, org.apache.sling.api.SlingHttpServletRequest request, org.apache.sling.api.SlingHttpServletResponse response, @NotNull org.apache.sling.api.resource.Resource rootResource, int limitRagTexts)	Answer a question with RAG from the given resources, e.g.
List<String>	searchRelated(org.apache.sling.api.resource.Resource root, String querytext, int limit)	Returns a list of up to limit paths that might be related to the query.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

markdownService

protected ApproximateMarkdownService markdownService

embeddingService

protected GPTEmbeddingService embeddingService

aiConfigurationService

protected AIConfigurationService aiConfigurationService

chatCompletionService

protected GPTChatCompletionService chatCompletionService

pageCachedValueService

protected PageCachedValueService pageCachedValueService

requestCounter

protected final AtomicLong requestCounter

Constructor Detail

RAGServiceImpl

public RAGServiceImpl()

Method Detail

searchRelated

@Nonnull
public List<String> searchRelated(@Nullable
                                  org.apache.sling.api.resource.Resource root,
                                  @Nullable
                                  String querytext,
                                  int limit)

Description copied from interface: RAGService

Returns a list of up to limit paths that might be related to the query. We search for the whole query, andy quoted strings inside and then for all single words in the query, scored. The paths are returned in descending order of lucene score.

Specified by:

searchRelated in interface RAGService

Parameters:

root - the root resource to search in

querytext - the query text

limit - the maximum number of paths to return

Returns:

a list of paths to jcr:content nodes, empty if no results or any of the parameters don't fit.

containsQuery

@NotNull
protected @NotNull List<String> containsQuery(@NotNull
                                              @NotNull org.apache.sling.api.resource.Resource root,
                                              @NotNull
                                              @NotNull String querytext,
                                              int restOfLimit)
                                       throws javax.jcr.RepositoryException

Throws:

javax.jcr.RepositoryException

normalize

@Nonnull
protected String normalize(@Nonnull
                           String querytext)

Turn it into a query for the words mentioned in there - that is, remove all meta characters for CONTAINS queries: AND, OR, words prefixed with -, quotes, backslashes. We use an OR query to find pages with as many words as possible.

orderByEmbedding

@Nonnull
public List<org.apache.sling.api.resource.Resource> orderByEmbedding(@Nullable
                                                                     String querytext,
                                                                     @Nonnull
                                                                     List<org.apache.sling.api.resource.Resource> resources,
                                                                     @NotNull
                                                                     @NotNull org.apache.sling.api.SlingHttpServletRequest request,
                                                                     @NotNull
                                                                     @NotNull org.apache.sling.api.SlingHttpServletResponse response,
                                                                     @NotNull
                                                                     @NotNull org.apache.sling.api.resource.Resource rootResource)

Finds the resources whose markdown approximation has embeddings that are the most similar to the querytext embedding. Useable e.g. as filter after searchRelated(Resource, String, int).

Specified by:

orderByEmbedding in interface RAGService

Parameters:

querytext - the query text

resources - the list of resources to search in

request - the request to use when determining the markdown approximation - not modified

response - the response to use when determining the markdown approximation - not modified

rootResource - the root resource to search in

ragAnswer

public String ragAnswer(@Nullable
                        String querytext,
                        @Nonnull
                        List<org.apache.sling.api.resource.Resource> resources,
                        @Nonnull
                        org.apache.sling.api.SlingHttpServletRequest request,
                        @Nonnull
                        org.apache.sling.api.SlingHttpServletResponse response,
                        @NotNull
                        @NotNull org.apache.sling.api.resource.Resource rootResource,
                        int limitRagTexts)

Answer a question with RAG from the given resources, e.g. found with searchRelated(Resource, String, int).

Specified by:

ragAnswer in interface RAGService

Parameters:

querytext - the query text

resources - the list of resources to answer from

request - the request to use when determining the markdown approximation - not modified

response - the response to use when determining the markdown approximation - not modified

rootResource - the root resource to find GPT configuration from

limitRagTexts - the maximum number of RAG texts to consider

Returns:

the answer text

getEmbeddingsCache

protected GPTEmbeddingService.EmbeddingsCache getEmbeddingsCache(Map<String,org.apache.sling.api.resource.Resource> textToResource)

collectSearchKeywords

@Nonnull
public List<String> collectSearchKeywords(@Nullable
                                          String querytext,
                                          @Nonnull
                                          org.apache.sling.api.resource.Resource rootResource)
                                   throws javax.jcr.RepositoryException

Processes a query to have the AI suggest a couple of search keywords for use with the other methods that might find the most relevant results.

Specified by:

collectSearchKeywords in interface RAGService

Parameters:

querytext - the query text for which we find keywords

rootResource - the root resource to find GPT configuration from

Returns:

a list of keywords

Throws:

javax.jcr.RepositoryException