Skip to content

Comparison

Should your team own this?

A real question with two real answers. Every team that wants grounded answers over its own content ends up building the same seven things — here is an honest inventory of them, and of what you give up by not owning them.

The inventory

What you would own.

Not the naive version of each — the version that survives a second source system and a bad week.

  1. A chunker

    Boundaries that do not cut a claim in half, and a strategy you can change later without re-embedding everything.

  2. An embedder

    A model choice you are stuck with, because changing it means recomputing every vector you have ever stored.

  3. A vector store

    Sizing, sharding, and a rebuild path for the day it loses a collection.

  4. An entity and relationship extractor

    Prompt design per content type, shard-level retries, and a cost ceiling so one tenant cannot spend your whole model budget.

  5. A graph store

    Entity resolution that holds under concurrent writers, so two workers processing two documents do not create two nodes for one thing.

  6. An ontology manager

    Versioning, and an apply path that does not destroy rows the new schema rejects — otherwise every vocabulary change is irreversible.

  7. An answer synthesiser

    Grounding enforced in the prompt, citation markers validated against the documents actually returned, and a refusal path when the corpus does not answer.

Plus the part that is not a system and takes the longest: orchestration that retries at the step rather than the job, bounds concurrency per tenant, and can replay a run that died halfway.

Why orchestration is the hard part

These are the numbers before and after we got it right.

Not a competitive benchmark — a record of what a naive pipeline does, and what fixing it is worth. This is the work you would be doing.

Documents lost
9 in 205improved tozero

Across the same corpus, after moving orchestration onto durable steps with per-step retries.

Extraction p90
~103 simproved to5.2 s

Per document, sharded and run concurrently under a per-organisation ceiling.

Throughput
3.0 / minimproved to17.1 / min

Documents fully processed per minute, same corpus, same hardware shape.

Worst pipeline step
348% of budgetimproved to4%

As a share of the load balancer timeout. Above 100% the request dies before it finishes.

The first version of this pipeline lost nine documents in two hundred and five, and its worst step took three and a half times the load balancer's patience. Neither failure was visible from the outside — documents simply were not there.

The honest trade

What you give up by not building it.

Buying is not free either. These are the four things you would be handing over.

What building gives you

What we do instead

Model choice. Your own extraction and embedding models, your own provider, your own region.

We select the models and pay for them. You cannot swap them, and that is a real constraint rather than a setting.

Deployment control. Your VPC, your keys, your compliance posture.

One managed deployment, one region. No bring-your-own-cloud and no customer-managed keys today.

A prompt you can tune per use case.

A synthesis prompt we tune, with grounding and refusal enforced. You can inspect every input to it, but not rewrite it.

No dependency on anyone else's roadmap.

A REST integration with no code-level coupling — but our roadmap is still ours.

Deciding

Which side you are on.

Build

Retrieval quality is your product's core differentiator
Included
Retrieval is a feature inside a product about something else
Not included
You must run in your own cloud or a specific region
Included
You need a specific model or provider
Included
You have an engineer who can own a pipeline indefinitely
Included
You want the first cited answer this week
Not included
You want a governable vocabulary without building versioning
Not included
Your content spans two or more source systems
Not included

Use GRaaS

Retrieval quality is your product's core differentiator
Not included
Retrieval is a feature inside a product about something else
Included
You must run in your own cloud or a specific region
Not included
You need a specific model or provider
Not included
You have an engineer who can own a pipeline indefinitely
Not included
You want the first cited answer this week
Included
You want a governable vocabulary without building versioning
Included
Your content spans two or more source systems
Included

FAQ

Frequently asked

Try the shortcut before you commit to the long way.

The free tier costs nothing and takes one POST. If it does not do what you need, you have lost an afternoon rather than a quarter.

  • no card required
  • first workspace is free
  • one POST to ingest