Soraya Zheng

Contributing Writer, Explore Agentic

About Soraya

Soraya writes the enterprise RAG pillar: chunk strategies including late chunking, the embedding-model landscape (OpenAI text-embedding-3, Cohere Embed v4, Voyage 3 and 4), re-rankers, RAGAS-style evaluation, vector-store selection, and the point at which agentic RAG earns its seat. The evaluation half of the beat runs through the agent-testing work — trajectory checks, LLM-as-judge rubrics, and CI gating that blocks a merge on a recall regression instead of a hunch.

Soraya also covers where retrieval meets the buying decision: MCP versus RAG as two protocols for two different problems, the eight-way Glean alternatives comparison sorted by procurement profile rather than feature grid, Claude Team versus Enterprise seats against Bedrock usage, and Moveworks after the ServiceNow acquisition. Reviews the RAG evaluation, LLM-as-a-judge, model-routing, and document-processing pieces. Expect named models, measured metrics, and an explicit note whenever a number originates in a vendor's own eval.