Vector Vs Graph Retrieval Selector

Vector-vs-Graph Retrieval Selector

Overview

Ch1 makes the vector-vs-graph choice evidence-based rather than ideological. Microsoft's BenchmarkQED classifies queries on two axes:

• Scope — local (specific facts in a small number of regions) vs global / sensemaking (reasoning over large portions of the dataset).
• Type — data (direct fact retrieval) vs activity (interpretive / strategic).

The chapter's numeric anchors:

• Vector RAG: ~90% accuracy on simple lookups (DataLocal); 20-30% on complex reasoning (ActivityGlobal). "The very mechanism that makes vector search efficient becomes its fundamental limitation."
• LazyGraphRAG outperforms vector RAG by 50-60% on multi-hop reasoning.
• EyeLevel.ai: at 100,000 pages, vector accuracy drops up to 12% while graph drops only ~2%.
• The larger-context-window rebuttal: BenchmarkQED tested vector RAG against LazyGraphRAG with a ~1-million-token window (essentially the whole dataset); vector RAG still lost on every query type except the most basic factual questions, and bigger windows worsen "lost in the middle."

Where vector RAG shines (Ch1): local, fact-based lookups — customer support, FAQ, recommendation. Where it collapses: multi-hop reasoning, temporal awareness, the associativity gap ("which services were affected by the database migration that followed the security patch we discussed last month").

Ch1's own recommendation for agents is a HYBRID architecture — parallel vector + graph (vector search -> graph traversal -> context synthesis) — because agentic behavior "requires constantly moving between local and global understanding." GraphRAG is not free: the chapter names upfront graph-construction cost, query latency that grows with graph size, contextual nuance lost in triples, and schema-evolution cost. The selector surfaces those costs whenever it recommends GRAPH or HYBRID.

When to Use

• Choosing a retrieval architecture for a new enterprise agent
• Answering "should we add a graph, or is vector RAG enough?"
• Rebutting "let's just use a bigger context window instead of a graph"
• Mapping a mixed query workload to the right per-query strategy
• Teaching the BenchmarkQED local/global x data/activity quadrants

Phrases: "vector or graph", "do we need GraphRAG", "will a bigger context window fix it", "retrieval architecture", "why does RAG fail on this query", "local vs global queries".

When NOT to Use

• Tuning an existing pipeline (chunk size, embedding model, reranker) — this chooses the architecture, not its hyperparameters.
• Consumer FAQ / support bots where DataLocal lookups dominate — the answer is VECTOR and you already know it.
• As a graph builder. This recommends graph; it does not construct one. See Ch3 skills (graph-model-selector, schema-pattern-selector).

Process

Rationalizations

Red Flags

• A multi-hop / temporal / global workload recommended VECTOR. The scope or multi-hop flags are unset; re-check the workload description against the associativity-gap example.
• GRAPH recommended for a latency-critical DataLocal lookup. Graph traversal is slower than an ANN lookup here; the selector should have chosen VECTOR or HYBRID — verify the flags.
• HYBRID chosen for everything. Either the workload is genuinely agentic (legitimate — that is the book's default) or the signals are under-specified; add concrete scope/type/multi-hop values.
• A GRAPH/HYBRID recommendation adopted without reading graphrag_costs. The upfront-construction and maintenance cost is real; budget for it before committing.

Non-Negotiable Verification

1. Run the benchmark battery. python cli.py benchmark must report 8/8:
   • DataLocal lookup -> VECTOR; ActivityGlobal multi-hop -> GRAPH; agentic -> HYBRID; graph-signals + unstructured -> HYBRID
   • the larger-context-window flag attaches the 1M-token rebuttal
   • 100k+ pages attaches the EyeLevel 12%-vs-2% scale note
   • GRAPH/HYBRID surface GraphRAG costs; VECTOR does not
   • the ActivityGlobal quadrant carries the 20-30% anchor
2. Verify CLI help. python cli.py --help exits 0 and prints the SKILL.md description.
3. Inspect the scenario. python cli.py scenario devops should recommend VECTOR for the 5xx lookup, GRAPH for the cascading-migration query, and HYBRID for the autonomous agent.

Security Posture

• Prompt injection. The selector consumes structured flags and booleans, not free-text documents, so there is no injection surface in lib.py. The numeric anchors and rebuttal text are author-controlled constants, not model-generated.
• Data exfiltration. No network calls; the only file read is the workloads JSON path the caller supplies (default: the bundled sample). --json output goes to stdout.
• Privilege escalation. No shell invocation, no dynamic import, no file writes. The recommendation is advisory; it selects an architecture and does not touch any datastore or credential.
• Decision integrity. Treat the recommendation as a starting point, not a mandate — the chapter's numbers are reported so a human can audit the rationale rather than trust an opaque verdict.

Source Attribution

Distilled from Agentic GraphRAG (O'Reilly, by Anthony Alcaraz and Sam Julien), Chapter 1 — Defining Agentic AI, "The Limitations of Vector-Based Retrieval" and "GraphRAG" sections. The BenchmarkQED quadrants, the 90% / 20-30% vector-RAG numbers, the LazyGraphRAG +50-60% multi-hop figure, the EyeLevel.ai 12%-vs-2%-at-100k-pages result, and the ~1-million-token larger-context-window rebuttal are all the chapter's, anchored in Microsoft's "From Local to Global: A GraphRAG Approach to Query-Focused Summarization" and BenchmarkQED research.