MindLox

AI & Retrieval

AI Knowledge Assistant

An internal assistant that answers from a company's own documentation, cites every source, and says plainly when it does not know.
AI & RetrievalConcept build
Discuss something like this

The problem

What this kind of product has to solve

Organizations accumulate knowledge in places search cannot reach: policy documents, onboarding guides, past project write-ups, support macros, and a wiki three reorganizations out of date. New staff ask colleagues because asking the system does not work.

The naive fix - connecting a language model to a document store - produces something worse than useless: fluent answers assembled from the wrong documents, with no way for the reader to tell.

The solution concept

How we built it

This concept build demonstrates a retrieval layer designed for verification rather than just fluency. Every answer is assembled from retrieved passages and rendered with inline citations that link back to the exact source section, so the reader can check the claim in one click.

Hybrid retrieval combines keyword and semantic search, because the two fail in different ways: semantic search misses exact identifiers like policy numbers, keyword search misses paraphrase. Results are re-ranked before they reach the model.

When retrieval confidence is low, the assistant declines and offers the closest documents instead of guessing. That behavior is scored in the evaluation set alongside correctness, because a system that is wrong confidently is more damaging than one that is occasionally unhelpful.

Engineering decisions

The choices that actually determine whether it works

Each of these could have gone the other way. The reasoning is the part worth reading - it is how we would argue the same case on your project.
  1. 01

    Citations are mandatory, not decorative

    The interface renders an answer only when it can attach sources. An uncitable answer is treated as a retrieval failure, which keeps the system honest about its own coverage.

  2. 02

    Evaluation set written before the interface

    A graded set of real questions with expected answers and expected sources. Every retrieval or prompt change is scored against it, so quality is tracked rather than debated.

  3. 03

    Permissions enforced at retrieval time

    Documents are filtered by the asking user's access before they reach the model. Access control that only applies to the interface is not access control at all.

  4. 04

    Cost routed by task difficulty

    Classification and reformulation use a small model; synthesis uses a larger one. Most requests never touch the expensive path.

Intended business value

What this is designed to achieve

  • Intended to reduce the time staff spend locating information that already exists internally, and the colleague interruptions that substitute for it.
  • Designed so that answers can be verified in one click, which is the condition under which teams actually trust and adopt an internal assistant.
  • Built to make quality measurable, so the system can be improved deliberately rather than adjusted on anecdote.

No outcome figures are shown because this is a concept build with no client deployment behind it. Quoting a percentage here would be fabrication, so we quote nothing.

Built with

  • Next.js
  • TypeScript
  • Claude API
  • PostgreSQL
  • pgvector
  • Hybrid BM25 + vector search
  • Streaming responses

Interface rendered from styled markup rather than a screenshot - crisp at any density, and no image request. Values shown are illustrative.

Related

Services and sectors this relates to

Your next big idea deserves better technology.

Tell us what you are building. Let's explore how MindLox can help bring it to life.

Start a ConversationSee how we work

Prefer email? hello@mindlox.io

Start a ProjectEmail hello@mindlox.io