AI-Powered Knowledge Base from Cisco Webex Meetings

The integration connects to the Cisco Webex API to access meeting recordings, transcripts, and shared files (PDFs, presentations) stored in Webex Cloud. A scheduled ingestion job pulls new content, processes the audio/video through a transcription service, and extracts text from documents. The core AI pipeline then performs topic modeling and entity extraction to identify key subjects, decisions, action items, and project names mentioned in the conversation. This structured metadata is used to auto-tag each meeting artifact.

The processed knowledge is formatted into a standardized article template and pushed via REST API to your chosen knowledge base platform—Confluence pages, SharePoint lists, or a custom wiki. The system creates a hierarchical structure: a parent page for the project or recurring meeting series, with child pages for individual sessions. Each page includes the transcript, extracted decisions/actions, linked files, and generated summaries. A vector embedding of the full content is stored separately to power semantic search, allowing users to ask questions like "What was decided about the Q3 launch?" across hundreds of past meetings.

Rollout starts with a pilot team and a defined set of meeting series (e.g., weekly product reviews, client check-ins). Governance is critical: we implement RBAC controls to respect existing Webex meeting permissions, ensuring private discussions aren't exposed. An approval workflow can be added where a meeting owner or project lead reviews the auto-generated article before it's published. Audit logs track all accesses and changes. The final step is integrating the new knowledge base into daily workflows—linking it from project management tools, training support teams to use it for faster answers, and setting up alerts when relevant past decisions are cited in new meetings.

The integration connects to the Cisco Webex API to access meeting recordings, transcripts, and shared files (like presentations or documents). A scheduled ingestion job, typically running as a serverless function or containerized service, polls for new meetings based on criteria such as calendar tags, specific hosts, or concluded meeting status. For each qualifying meeting, the system retrieves the transcript (via the recordings and transcripts endpoints) and any associated content from the meeting/files endpoint, storing raw payloads in a secure object store like AWS S3 or Azure Blob Storage for auditability and reprocessing.

The core AI processing pipeline then executes: 1) Chunking & Embedding: Transcripts are split into logical segments (e.g., by agenda item or speaker turn). Each segment, along with extracted text from shared documents, is converted into vector embeddings using a model like OpenAI's text-embedding-3-small and indexed in a vector database such as Pinecone or Weaviate. 2) Summarization & Tagging: A separate LLM call (e.g., to GPT-4 or Claude 3) analyzes the full meeting context to generate a concise summary, extract key decisions, and assign relevant topic tags (e.g., project_alpha, budget_review, q3_roadmap). This structured metadata is stored alongside the vector pointers. 3) Article Generation: Using the summary and key segments as a source, a final LLM call drafts a formatted knowledge base article, following the template of the target system—Confluence or SharePoint.

The generated article, along with its metadata and vector ID references, is posted to the destination via its REST API (e.g., Confluence Cloud API or Microsoft Graph API). Governance is enforced through a human-in-the-loop approval queue for sensitive meetings before publishing, and all actions are logged with meeting IDs, user IDs, and timestamps for compliance. The final architecture enables users to semantically search the knowledge base ("Show me discussions about migrating off Server 2012") and be directed to the exact moment in the source recording, creating a closed-loop system where meetings become a proactive organizational asset.

The integration architecture treats the Cisco Webex API as a secure event source. Meeting transcripts and shared files (from Webex recordings and meetingContents APIs) are ingested into a processing queue. Here, access controls are enforced: only meetings from authorized spaces or with specific meetingType tags are processed. Sensitive content can be filtered using regex patterns or keyword lists before any AI model interaction, ensuring compliance with internal data policies. The processed content is then passed through a RAG (Retrieval-Augmented Generation) pipeline where a language model, grounded in your existing Confluence or SharePoint taxonomy, generates structured draft articles.

A human-in-the-loop approval step is critical before publication. Draft articles are posted to a designated staging area in Confluence (e.g., a Draft KB space) or a SharePoint library with a Pending Review metadata flag. Designated knowledge managers or project leads receive notifications via Webex or email to review, edit, and approve. This governance layer ensures accuracy, maintains editorial voice, and prevents the automatic publication of incomplete or sensitive discussions. The system maintains a full audit trail linking the source meeting, the processing job ID, the reviewer, and the final published article.

We recommend a phased rollout starting with a pilot team. Phase 1 targets a single project team's recurring strategy meetings, automatically creating knowledge articles in a sandbox Confluence space. This validates accuracy, refines tagging logic, and socializes the workflow. Phase 2 expands to department-level all-hands and project review meetings, integrating with official project spaces in SharePoint. The final phase rolls out org-wide, with configurable rules per meeting series—for example, executive briefings may only generate private summaries, while engineering deep-dives auto-publish to public tech hubs. This controlled approach maximizes adoption while managing risk and content quality.