Synthesise — RAG · GraphRAG · Feature mesh
Synthesise fuses vectors, graphs and features into a single retrieval plane your agents can actually trust. Every chunk carries policy, every feature has a point-in-time guarantee, every answer is evaluated.
Retrieval p50
<80ms
Answer faithfulness
>0.92
Feature freshness
<5s online
Deliverables
Everything that ships
- Vector store (pgvector / Pinecone)Chunked, embedded, governed corpora with TTL + ACL.
- Knowledge graphEntities, relations, temporal edges — queryable via Cypher / GraphQL.
- Online feature storeFeast / Tecton — point-in-time correct features for serving.
- Hybrid retrieverBM25 + dense + graph re-rank, evaluated on golden Q&A set.
- Eval harnessRAGAS + custom metrics, gated by CI for every embedding change.
Pod composition
- ML Engineer
- Knowledge Engineer
- Backend Engineer
Example output · Retriever call · /v1/retrievejson
{
"query": "Why did churn spike in EU-West last week?",
"scope": ["crm", "billing", "support"],
"k": 8,
"rerank": "graph",
"policy": "agent.read.governed"
}Timeline
Weeks 4–9 · retrieval plane live by day 63
- 1Weeks 4–5Corpora + embeddings
Govern, chunk, embed; TTL + ACL on every vector.
- 2Weeks 5–7Graph + features
Entity/relation graph + online feature store stood up.
- 3Weeks 7–9Hybrid retriever + evals
BM25 + dense + graph rerank, RAGAS gating in CI.
FAQs
Things prospects ask
Self-host or managed?
Both. Default is pgvector + a small graph DB on your cloud; we'll run Pinecone/Neo4j Aura if you prefer managed.
How do you stop hallucinations?
Hybrid retrieval, citation-required prompts, faithfulness scoring and a refusal policy when score < threshold.
Commission · S4 Synthesise
Stand up Synthesise in Weeks 4–9.
We'll respond within one business day with a scoping note, a fixed-price outcome contract, and a named principal. Your details sync straight into our concierge queue.
- • Outcome-priced — no T&M.
- • Sovereign by default — your data, your region, your keys.
- • Wired into the Fuel Pressure gauge from day one.