Embedding Store Selection Gate (Onboarding)

Embedding Store Selection Gate (Onboarding)

1. 01

   Classify the client workload before any vendor demo is booked

   Separate the request into semantic search, RAG retrieval, recommendation, or hybrid analytics. A knowledge assistant for 40,000 support documents and a product recommender have different index shapes, refresh cadences, and filter requirements.

2. 02

   Size the corpus in vectors, not documents

   Multiply document count by chunk count per document and record the embedding dimension. A 200-page policy library chunked at 500 tokens can land near 60,000 vectors, which changes whether a managed tier or a self-hosted node is the right starting point.

3. 03

   Write down the query latency budget the client experience actually needs

   Sub-50ms targets suit customer-facing search boxes, while overnight batch retrieval for internal reporting tolerates seconds. Put the number in the statement of work so delivery is measured against it.

4. 04

   Map the filter and hybrid requirements to the shortlist

   Confirm whether the client needs metadata filtering, dense-sparse hybrid retrieval, or full-text fallback. Qdrant and Zilliz both publish hybrid query support, while Meilisearch leads with typo-tolerant full-text plus semantic layers, so match the engine to the dominant query pattern.

5. 05

   Decide hosted versus self-managed against the client's ops reality

   A managed platform such as Pinecone or Zilliz removes index rebalancing work; a self-hosted Weaviate or Qdrant deployment trades that convenience for infrastructure the client's team must staff. Ask who holds the pager at 2am before recommending either.

6. 06

   Score lock-in exposure on a one-page matrix

   Record export formats, index portability, and pricing meter ownership for each candidate. Forrester's November 2026 research flags vendors that set usage limits, control feature access, and arbitrate their own disputes, so note which shortlisted services hold all three roles.

7. 07

   Run a two-day proof of concept on the client's own data

   Load a representative slice, measure recall on 20 known-good queries, and log latency at the 95th percentile. Label the output a PoC, not a pilot, so scope expectations stay aligned with the deliverable.

8. 08

   Sign off the selection memo with the client's technical owner

   Attach the latency budget, cost estimate at 12-month volume, and exit plan. Retainer work on AI features stalls when the storage decision is made in delivery rather than onboarding.