System 05 · Retrieval

Agentic RAG systems

An agentic assistant that answers from your own material instead of guessing. It cites what it used, and it says it does not know rather than inventing something plausible.

RetrievalEmbedded, over your corpus
GroundingAnswers cite their source
RefusalSays "not in the material"
HostingOur cloud, or yours
Architecture

Three layers, so the guessing stays in one of them.

Directive

What to do

The SOP, written down. Goals, inputs, the tools to use, the outputs and the edge cases — in plain language, versioned like code. When we learn an API limit the hard way it gets written here, so it is never learned twice.

Orchestration

Decision-making

The only probabilistic layer. It reads the directive, calls the tools in order, handles the errors and asks when something is genuinely ambiguous. It does not do the work itself.

Execution

Doing the work

Deterministic scripts. Same input, same output, every run. Ninety per cent accuracy per step is fifty-nine per cent over five, so anything that must be right lives down here instead.

Data flow

In, through, out.

Sources
Your documents, pages and knowledge base
Product and pricing material
Past support and sales conversations
Process
  1. Chunk and embed the corpus
  2. Retrieve the passages that actually match
  3. Answer strictly from what was retrieved
  4. Attach citations to every claim
  5. Log the unanswered questions
Artifacts
Embedded, queryable index
Grounded answer with citations
Captured lead from the conversation
Gap report: what it could not answer

The gap report is the part clients end up valuing most. Every question the corpus could not answer is a content brief you did not have to guess at — the assistant tells you what your material is missing.

Integrations

It reads your stack. It doesn't replace it.

  • ClaudeGrounded answering
  • OpenAIEmbeddings + fallback
  • PineconeVector retrieval
  • QdrantSelf-hosted vectors
  • PostgreSQLSource of record
  • LangChainRetrieval orchestration

Also connects to

Your CMSGoogle DriveNotionClickUpSlackAnalytics 4

Read-only wherever read-only is enough. Nothing gets write access it doesn't need, and every write is logged.

Quality gates

Checked before it ever reaches you.

Every answer is traceable to retrieved passages — no ungrounded generation.

Below the confidence floor it refuses instead of guessing.

Unanswered questions are logged rather than silently dropped.

That consistency is the whole point. The same checks and the same artifacts on every run, so when a number moves you know it moved because your site moved — not because the method did.

Datasheet

The boring details.

SpecStandardDeployable
Where it runsOur infrastructureYours
Who holds the API keysUs, scoped per clientYou
CadenceMonthly, or weeklyAny schedule
Raw data retentionRebuilt each runYour policy
Artifact deliveryYour Drive and trackerYour choice
Source code accessFull
Runs unattendedYesYes
Failure alertingSlack and emailYour channels
The other systems

Same architecture, different job.

Want it in your own stack?

Deployable hands over the software, the hosting and the keys. You run it; we stay on for support.

Talk scope
Before you start

Questions about RAG systems.

What is a RAG system?

A RAG system is an assistant that answers from your own documents instead of from general training data. Retrieval-augmented generation means it finds the relevant passages in your material first, then answers from them, and cites what it used.

What does it do when it does not know?

It says so. The system is built to decline rather than produce a plausible answer it cannot ground in your material, because a confident wrong answer costs more to undo than no answer costs to escalate.

Where does our data live?

Wherever you choose. On Deployable the system runs inside your own infrastructure with your own keys, so your documents never leave your environment.