Larger context windows have not removed the need for retrieval. Enterprise knowledge changes, lives in many systems and is governed by permissions that differ from user to user. Agents also need memory, but indiscriminate storage of previous conversations creates privacy, accuracy and contamination risks.
This course treats retrieval, context and memory as complementary parts of one information architecture. Participants will build systems that locate authoritative evidence, preserve provenance, maintain selected state across tasks and explain which sources support an answer or action. They will also test the complete retrieval-to-outcome path rather than relying only on generic similarity scores.
The course is framework-neutral and addresses documents, structured data, code, conversations and operational knowledge.
Learning Outcomes
Upon completion of this course, participants will be able to:
- Choose appropriately between retrieval, long context and persistent memory
- Design ingestion and indexing around authority, structure and permissions
- Implement lexical, semantic and hybrid retrieval
- Build agentic retrieval that plans and refines multi-step searches
- Separate task state, user memory and organisational knowledge
- Apply provenance, freshness and deletion controls
- Defend knowledge agents against untrusted or poisoned content
- Evaluate retrieval, grounded responses and end-to-end task outcomes
Course Outline
The Enterprise Knowledge Stack
- Long context, RAG, search and memory as complementary capabilities
- Types of enterprise knowledge and their operational characteristics
- Authoritative systems, derived indexes and temporary working context
- Reactive question answering versus tool-using knowledge agents
- Latency, quality, permission and cost trade-offs
- Selecting an architecture from the use case rather than the trend
Ingestion and Knowledge Preparation
- Connectors for documents, databases, code and collaboration systems
- Parsing structure, metadata, tables and attachments
- Chunking by semantic and document boundaries
- Preserving identifiers, versions and source locations
- Incremental ingestion, updates and deletion
- Handling duplicated, conflicting and low-quality content
Retrieval Design
- Lexical, dense, sparse and hybrid retrieval
- Metadata filtering and permission-aware queries
- Query rewriting and decomposition
- Re-ranking by relevance, authority and freshness
- Retrieving structured facts through tools and APIs
- Returning evidence in a form an agent can inspect and cite
Agentic Retrieval
- Planning searches across several sources
- Iterative retrieval when initial evidence is incomplete
- Tool selection for documents, code and structured systems
- Detecting contradictions and requesting clarification
- Limiting search depth, loops and unnecessary model calls
- Recording the evidence path for later review
Grounded Generation and Action
- Separating retrieved content from system instructions
- Requiring claims to map to evidence
- Handling insufficient or conflicting sources
- Calibrating uncertainty and refusal
- Using retrieved evidence to support tool actions
- Presenting citations that users can verify
Memory Architecture
- Working context, durable task state and persistent memory
- Episodic, semantic and procedural memory patterns
- User-specific versus shared organisational memory
- Selecting what is worth remembering
- Summarising without converting assumptions into facts
- Retrieval from memory alongside current authoritative sources
Memory Governance
- Consent, purpose limitation and data minimisation
- Isolation across users, teams and jurisdictions
- Correcting, expiring and deleting stored memory
- Preventing stale memory from overriding current evidence
- Audit records for memory creation and use
- Protecting sensitive information from prompts and traces
Security and Content Integrity
- Indirect prompt injection through retrieved content
- Source trust, provenance and content signing
- Poisoned indexes and compromised connectors
- Limiting available tools during untrusted-content processing
- Preventing data exfiltration through generated answers or actions
- Designing safe behaviour when detection controls fail
Evaluation and Operations
- Retrieval relevance, recall and ranking measures
- Groundedness, citation correctness and answer usefulness
- Scenario evaluation for multi-step knowledge tasks
- Permission and data-leakage tests
- Monitoring freshness, latency, cost and failed retrieval
- Turning production feedback into regression cases
Practical Capstone
- Design a knowledge agent over mixed enterprise sources
- Build permission-aware hybrid retrieval
- Add task state and governed persistent memory
- Require evidence for generated claims and actions
- Test stale, conflicting and malicious content
- Present evaluation results and an operational governance model
Intended audience
This course is designed for software developers, AI engineers, data engineers, knowledge-management specialists and solution architects building assistants or agents over enterprise information. It is suitable for teams moving beyond basic vector-search demonstrations towards reliable production knowledge systems.
Prerequisites
Those attending this course should meet the following:
- Proficiency in Python or another server-side language
- Familiarity with APIs, databases and structured data
- Basic understanding of embeddings and large language models
- Awareness of authentication and enterprise data permissions