Build a Company Knowledge Graph Your AI Can Use
EnrollAI works well for one person and badly for a team, because a team's knowledge is scattered across documents, threads and people's heads. A knowledge graph addresses that. It holds entities, relationships and sources that your agents can traverse, so answers cite what the organization actually knows rather than inventing it. This course builds one for a real team, which is yours, from extraction through retrieval to the agents that use it. The result is one durable data layer rather than another tool with a short life.
How does this prepare you for a job?
Search, platform and AI-infrastructure teams all converge on this shape of work, and structured knowledge plus retrieval appears as a named line in their job descriptions.
- Built a knowledge graph from unstructured company documents, using LLM extraction of entities and relationships, stored in Postgres so it can be queried.
- Combined graph traversal with vector retrieval, meaning KG plus RAG, to ground agent answers in internal sources that are cited.
- Designed the freshness pipeline covering change detection, re-extraction, and how conflicts are handled as documents evolve.
- Added retrieval that is aware of permissions, so answers respect the access controls on the underlying documents.
- Shipped a team Q&A agent over the graph with groundedness evals and citations that are mandatory rather than optional.
AI tools learnt
A pragmatic graph stack, staying relational until something proves it should not be.
Hands-on skills
- Extract entities and relations from messy documents with LLMs, and measure how good the extraction is
- Choose the store, deciding between a property graph and relational-plus-vectors on the basis of your query shapes
- Blend traversal with embedding retrieval so that multi-hop questions resolve
- Design for freshness, with change detection and re-extraction that do not corrupt the graph
- Respect permissions during retrieval, which is the requirement that kills naive team RAG
- Eval groundedness, so that answers either cite their sources or do not ship
Modules
Level: Basic
Model the team's knowledge
- The entities and relations that matter to your organization, starting with a light schema
- Extract from fifty real documents using Claude
- Measure extraction quality before you trust it
Store and query
- Setting up Postgres with pgvector, so traversal and similarity sit side by side
- The first queries that answer something a search box cannot
- Multi-hop questions, meaning the ones that need two edges to resolve
The team Q&A agent
- Wire graph retrieval into an agent, with citations required
- Groundedness checks run before rollout
- Ship v1 to your own team and watch what they ask it
Level: Advanced
Keep it alive
- Change detection and re-extraction pipelines
- Handling conflicts when two sources disagree
- Graph rot, meaning how to spot it and what to prune
Permissions and trust
- ACL-aware retrieval, tested adversarially rather than assumed
- Audit, covering who asked what and which sources the answer came from
- The groundedness eval, run as a standing gate
Org-scale lessons from the big graphs
- What Google's Knowledge Graph teaches about entity resolution and authority
- Why search-infrastructure teams are preoccupied with freshness, and what that implies for yours
- Write the design doc for your graph as though it served an organization of ten thousand people
Students can take the full course (both levels) or either level on its own.
Register
Waitlist
Join the course waitlist
No cohorts are scheduled yet. Join the waitlist and we’ll email you when registration opens.