Skip to content

Platform/Company Brain

Your systems get profiled where they sit, and the profile is kept honest.

An agent is only as correct as the slice of context it was handed. This is what decides that slice, and it holds the meaning of your data rather than a copy of it.

Profiled, not copied Agents read your schemas, samples, documents and APIs, and build the maps. Statistics and column roles come to us; the rows are queried live at answer time.
Refreshed on a schedule A brain built once starts going stale the same week. Profiling re-runs, diffs itself against what it believed, and applies what it is confident about.
Unsure is a state it can be in Anything ambiguous stops and becomes a review item with a name on it, instead of a confident answer nobody can trace.

What actually moves

The rows never leave your systems.

What SynOS holds is the statistics, the semantics and the bindings. The answer is assembled by querying your systems at the moment the question is asked.

Fig. 02 · the same table under both approaches

What it holds

Eight kinds of memory, because an agent needs more than documents.

Most context tools give you one shape, chunks of text with embeddings, which covers a policy document and none of the rest.

Fig. 09 · one record of each kind, and the store it lands in

Status is per capability and stated the same way everywhere on this site.

Getting the right slice

The long context window made retrieval more important, not less.

Most answers that get blamed on the model are a retrieval problem. A router picks the mode per query.

mode

what it does

when it fires

Auto routermode = auto

A selector picks the mode and the source per query.

the default, mixed query types over mixed sources

LexicalBM25

Sparse retrieval on exact term overlap. Fast, transparent, and it needs no GPU.

order IDs, account numbers, SKUs, policy codes

Semanticdense embeddings

Ranks by meaning, so paraphrase, synonyms and a second language all still land on the right passage.

natural language questions, varied vocabulary

Hybrid with rerankBM25 + vector + RRF

Runs lexical and semantic in parallel, fuses the two rankings, then a cross encoder reorders the top of the list for precision.

production default, prose mixed with literal identifiers

Agenticplan, retrieve, reason

Decomposes a hard question into sub questions, runs each, merges and reranks what came back.

multi part questions, long threads, follow ups

Graph traversalentity hops

Walks linked entities and their relationships. Traversal happens at query time, so there is no giant persisted graph to keep in sync.

multi hop reasoning across connected entities

Whatever the mode, what comes back is a typed context pack: scoped to the caller's role and project, ranked by authority and freshness, fitted to a token budget, and carrying citation IDs so an answer can be traced to the row it came from.

The part that decays

Schemas drift and definitions change. Agents keep answering anyway.

The failure nobody demos, because it takes a quarter to appear: the agent carries on sounding exactly as confident as it did in week one.

01

Re-profile

Agents re-read schemas, samples, documents and APIs. New tables, dropped columns, renamed fields, distributions that have shifted.

02

Diff and score

Compare what is there now against what the brain believed. Each mapping gets a confidence score and the date it was last confirmed.

03

Apply or escalate

High confidence changes apply themselves. Anything ambiguous stops and asks a person, which is the whole point of scoring it.

admin review queue

is the brain still right

STALE

7 tables not re-profiled in 30 days, and 2 of them feed a live agent.

DRIFT

orders.status gained 3 enum values, so the existing mapping no longer covers them.

CONFLICT

Two tables both claim revenue. Which one is authoritative?

LOW CONF

4 metric bindings sit below threshold, which means the agent is guessing on them.

PROPOSED

Profiling found a new entity, Gift Recipient, spanning 3 tables. Add it to the ontology?

Illustrative queue. Your systems produce your own.

What it is and is not

Three things people expect it to be, and what it does instead.

Not a warehouse It holds no copy of your operational rows and it is not where you run analytics. It knows which warehouse to ask and how to ask it.
Not an index you load There is no bulk load step to run before it works and no re-load when your data changes. Profiling reads your systems and refreshes the maps.
Not static retrieval Its content changes from use. Corrections, promoted rules and re-profiling results all land back in it, which is what the learning loop is for.

Bring one system and see what the profile finds.

The fastest version of this conversation is a walkthrough against a schema you already know well, because you will spot immediately where a profile is right and where it is guessing.