Custom RAG Development for Internal Company Knowledge

Neaptide builds RAG systems that help staff find and use information in company documents. We start with a limited knowledge set, named user roles and questions the system must answer or decline. The proposal defines the data sources, acceptance checks and handover before implementation begins.

When this service fits

A custom assistant is worth assessing when staff need answers across approved documents, ordinary search leaves too much manual work or the workflow has access rules an off-the-shelf tool cannot meet. Typical starting points are internal procedures, support knowledge and product reference material. If existing search solves the task, we will scope that simpler route instead.

RAG retrieves relevant documents and supplies them to a language model. It can make an answer easier to check, but citations alone do not guarantee accuracy. We plan tests for missing, outdated, conflicting and restricted information alongside ordinary questions.

Define the document boundary first

We inspect representative documents before promising a connector or parsing method. Scans, tables, conflicting versions and fragmented permissions can change the work substantially. Access and hosting decisions belong in the scope; a claim to support “all company data” would hide the choices that matter.

From a working prototype to an acceptance decision

The first build uses the agreed sources and a small set of representative questions. We inspect retrieval as well as the final response. If an answer is wrong because the current document was never retrieved, that needs a different correction from a model ignoring evidence it already received.

• Show a source reference the authorised user can open.
• Check an ordinary user and a restricted role against the same question.
• Test a missing answer and an obsolete procedure.
• Record expected behaviour, actual output and the evidence for every acceptance case.
• Agree which failures block release and which require a human review path.

Our published retrieval case shows a successful recovery of a real Neaptide service-price passage, with its source URL and surrounding terms. It is a selected observation, not an overall benchmark. Private permissions and answer generation need separate evidence in the actual project environment.

• Read our retrieval experiment and its limits
• Selected successful retrieval trace · JSON

Agree the handover you need

Before the build, we record which application code, configuration, evaluation files, ingestion jobs and operating instructions belong in the handover. The agreement also identifies components supplied under third-party licences or hosted subscriptions. These terms are project-specific; model access and a vendor platform are not assets we can promise to transfer as our own code.

The operating walkthrough covers source updates, access changes, failed requests and the route for reporting incorrect answers. If your team will maintain the application, include a reproducible deployment and a runbook check in acceptance. Continuing support is agreed separately with a named owner and scope.

How we estimate the work

The quote depends on source count and quality, access complexity, integrations, hosting constraints and the evaluation coverage. A prototype over clean exports is a different engagement from a production assistant with live connectors and permission changes. We agree the first milestone and the decision to continue before expanding the build. There is no universal RAG tariff or guaranteed accuracy percentage on this page.

• AI implementation cost categories
• Compare RAG providers and handover questions

Bring a small, safe example

To assess fit, describe the users, the documents they need and one recurring question. Redacted samples and a role list are enough for an initial discussion. Keep confidential documents in your approved environment until access and handling arrangements are agreed.