Knowledge plane / Tier 1 / Phase 1
Retrieval & Context Engineering
What information matters now?
Assemble the smallest sufficient context for each step, inside a token budget: hybrid retrieval, reranking, graph traversal, iterative refinement, compaction and tool search.
- Hybrid retrieval (lexical plus dense) with reranking
- GraphRAG and structured retrieval where relationships matter
- Agentic, iterative retrieval that refines its own queries
- Context pruning and compaction to fight bloat
- Token budgeting per step and per task
- Tool search: retrieving relevant tool definitions instead of loading all of them
governed corpus and entitlements
recalled memory joins the context
query grounded in agreed terms
delivers the assembled context to the model
context tokens are a first-class cost
groundedness and recall graded
retrieved content is untrusted input
AI Engineering, with Data Engineering
- Head of AI Engineering — accountable owner
- Retrieval / information-retrieval engineer
- Context engineer — budgets, compaction, prompt assembly
- ML engineer — embedding and reranker selection
- Data engineer — indexing pipelines and freshness
Architectural recommendation, not a research finding.
Pilot ships hybrid retrieval with reranking and a hard token budget. Add graph retrieval and iterative refinement once you can measure groundedness.
Signals to watch
- Recall at k and reranker gain
- Groundedness and citation accuracy
- Context tokens per successful task
- Retrieval latency P95
- Share of answers with a verifiable source
- Context bloat and lost-in-the-middle: more context, worse answers, higher cost
- Retrieval poisoning — hostile content in the corpus becomes instructions
- Tool definitions consuming the context window; tool search cuts that overhead by roughly 85%
- Index staleness invisible to the user
Tool and standards snapshot · last reviewed approximately mid-2026
Open source
- LlamaIndex
- Haystack
- Qdrant, Weaviate or Milvus
- pgvector
- OpenSearch or Elasticsearch hybrid search
- BM25 plus bge-reranker
- ColBERT
- RAGAS for retrieval evaluation
Managed
- Managed vector and hybrid search services
- Enterprise search platforms with connectors
Standards
- MCP tool-search patterns for tool-definition retrieval
- Structured citation formats carried into answers
Retrieval, memory and runtime state are three different things. Retrieval is what the system looks up now; memory is what it chose to keep; runtime state is where the task currently stands. Merging them is a common and expensive modelling error.