Skip to main content

Organization Context System

The Organization Context System gives Lira deep knowledge about your company — so she can give informed, relevant responses during meetings instead of generic AI answers.

Architecture​

The system is built across 9 phases:

Phase 1: Organization CRUD & Membership​

Basic organization management — create, update, delete orgs, invite members, manage roles.

Phase 2: Organization Profile & Context Injection​

Organization profile data (name, industry, description, values) is injected into Nova Sonic's system prompt at session start.

Phase 3: Website Crawl & Knowledge Base​

Lira crawls your company website using Crawlee + Cheerio, extracts text content, and indexes it for retrieval.

Phase 4: Document Upload, Parsing & RAG Pipeline​

Upload PDFs, DOCX, and text files. Content is extracted, chunked, and prepared for embedding:

  • pdf-parse for PDF extraction
  • mammoth for DOCX conversion
  • Custom chunking with overlap for context preservation

Phase 5: Vector Search & Semantic Retrieval​

Content chunks are embedded using OpenAI text-embedding-3-small (1536 dimensions) and stored in Qdrant (self-hosted on EC2):

Document → Chunks → OpenAI Embeddings → Qdrant Collection
                                              │
Meeting question → Embedding → Cosine Search ─┘
                                              │
                              Top-K relevant passages

Phase 6: Context Injection into Nova Sonic​

During a meeting, when Lira is about to respond:

  1. Recent transcript is embedded
  2. Vector search finds the most relevant knowledge base passages
  3. Passages are injected into Nova Sonic's context (token-aware, with caching)
  4. Mid-meeting context refresh keeps the context current

Phase 7: Task Execution Engine​

Tasks extracted from meetings go through an AI execution pipeline:

  1. Extraction — GPT-4o-mini identifies action items
  2. Review — Lira validates tasks (lira_review_status)
  3. Execution — Tasks are created in connected tools (Linear, GitHub, etc.)
  4. Notification — Webhook and Slack notifications confirm execution

Data Model​

interface Organization {
  orgId: string;
  name: string;
  industry: string;
  description: string;
  website: string;
  members: OrgMember[];
  integrations: Integration[];
  knowledgeBase: KnowledgeDocument[];
  usageQuota: UsageQuota;
}

Security & Data Isolation​

  • All knowledge base content is scoped to the organization
  • Vector search queries are filtered by orgId
  • No cross-organization data leakage
  • Documents are stored in S3 with organization-prefixed keys