Confluence RAG indexing and MCP integration

What OKC indexes in Confluence Cloud

The read-only connector uses the REST API to read published content in authorised spaces and pages. It keeps titles, sections, tables, versions and links. Incremental synchronisation picks up changes. You can select a visual analysis scope separately from the text indexing scope.

Text and tables

The index splits pages into searchable passages. Each passage keeps its section and version. Table rows remain associated with their headings.

Images and screenshots

OKC retrieves the pixels of authorised images. The model records visible text, objects, colours and relationships. For a workflow, it details written steps and roles, along with readable arrows, directions, labels and conditions. The description and these observations are indexed with each occurrence’s context. Keywords shown in the example are terms in this search text, without a separate tag filter.

Attachments

Supported attached images contribute to visual search. PDF, Word, Excel and PowerPoint documents are downloaded at their specific version and processed by the local readers shared by file sources and Confluence. Passages retain the filename, version and location. Coverage distinguishes extracted, pending, empty, unsupported and failed documents.

Federate Confluence and your file sources

The document module indexes PDFs, Word documents, Excel workbooks, PowerPoint presentations and text formats in authorised locations. Readers retain passages and their locations, such as a PDF page or Excel cell coordinates. Depending on the format and reader, OCR can extract text from a scanned document. Detailed visual analysis of every figure inside a PDF is a scope to assess separately.

Federated search queries the Confluence and file indexes, then combines their rankings. Each corpus retains its own representation model. Before returning evidence, OKC checks the Confluence source or original file. The interface offers both corpora; the API and MCP combine them with knowledge_search(sources=“all”). Unextracted content and errors remain visible in coverage reporting.

From a keyword to a source-linked answer

Lexical search finds a specific reference or exact button label. Semantic search uses embeddings to identify passages with similar meaning. Hybrid search combines the two to rank text, tables and image descriptions. The reference architecture uses PostgreSQL and pgvector.

RAG (Retrieval-Augmented Generation) supplies the model with passages selected for the question. OKC rechecks the sources in Confluence before and after generation. Readers receive an answer with citations and can open the relevant page. The reference installation runs its indexing, vision and generation models locally.

How we check quality with your team

Ovyka and your subject experts choose questions about a passage of text, a labelled screenshot and an image without text. For each case, we check separately whether the search finds the right attachment, whether its description captures the visible details and whether the citation points to the right page and version. Coverage reporting distinguishes images awaiting analysis from those already processed.

These checks guide adjustments to the scope, visual descriptions and retrieval before we connect the chosen channel. At handover, your administrators know where to monitor changes and when to ask a subject expert to check a result.

An API for your applications, MCP tools for your agents

Your portal can call the API to search, ask a question or open a document. An MCP-compatible assistant can also call the read-only tools directly. It receives passages and source links and can compose its reply with its own model.

Retrieval

knowledge_search ranks passages. knowledge_images searches images by their visual description and context.

Reading and coverage

knowledge_document opens the details of a page found by search. knowledge_status reports the scope and state of the index.

Files and combined search

knowledge_files searches documents by content, path or type. knowledge_file opens located passages after checking the original. knowledge_files_status reports document coverage. knowledge_search(sources=“all”) combines Confluence and files while keeping source references.

Skills and instructions

A skill tells the assistant when to call the tools, which evidence to retain and how to present it. Ovyka adapts these instructions to your channel, terminology and readership. Reading and sharing permissions apply independently of writing style.

How the API, MCP and a skill fit together

OKC API

Your application or portal calls search and reading functions directly. It controls the user flow and the presentation of the answer.

MCP server

A compatible assistant discovers the retrieval tools, calls them and receives passages with source links. The same adapter can serve several AI clients.

Assistant skill

Instructions tell the assistant when to search, which details to check and how to cite the source. They use a connected tool through MCP or a platform-specific action.

For several assistants, Ovyka recommends a common API and an MCP adapter, with instructions tailored to each client. For an Atlassian-only workflow, a Forge action can call the API and be used by a Rovo skill.

Choose a connection for your environment

Prepare an internal, partner or customer-facing integration

The local reference integration uses one reading identity. Before shared access, Ovyka defines the corpus for each audience, checks permissions per reader and tests revocation. An internal answer can cite a Confluence procedure. For an external customer, the assistant uses authorised sources and links to documentation the recipient can open.

Tell us which Confluence source, chat or portal you have in mind and how content may be shared. We can show you an architecture that fits that workflow.