Project / Caseworker

In development

Caseworker.
Evidence before answers.

An AI-assisted knowledge system, explored through source preparation, retrieval, grounded answers and evaluation.

CaseworkerIllustrative interface

⌕ Search▤ Documents▧ Collections

Find information

Search documents, policies and guidance ⌕

Source preparation

Clean, structured, traceable content

Evidence and context

Connect an answer to its source

Evaluation

Test relevance and missing information

FocusKnowledge retrieval

ContributionArchitecture & development

Evidence statusPublic artefacts in preparation

01 / The problem

Finding information is only the beginning.

When information is spread across documents, finding a plausible passage is not enough. The user needs relevant context, a traceable source and a clear indication when the available evidence cannot answer the question.

Caseworker explores that workflow. Its scope is a knowledge-assistance system, with access controls and evaluation treated as design concerns from the outset.

02 / The approach

Keep the route to the evidence visible.

1Prepare sourcesClean · chunk · describe

→

2Retrieve contextFilter · rank · inspect

→

3Draft & checkCite · evaluate · review

Proposed workflow. Retrieval and answer behaviour remain development and validation work.

Explore the architecture

03 / Design decisions

Make the trade-offs explicit.

Chunking: context versus precision

Small chunks can retrieve precise passages but lose surrounding meaning. Larger chunks retain context but may introduce irrelevant material. The choice needs testing with representative questions and source structures.

Retrieval: similarity versus relevance

Vector similarity is one signal. Metadata filters, access control, keyword matching and reranking may be needed. Evaluate whether the retrieved material actually supports an answer.

Answering: helpfulness versus unsupported claims

Ground the response in available evidence. Show citations and missing information, and define when the system should ask a clarifying question or decline to answer.

04 / Implementation

A build with visible stages.

Source preparation

Corpus preparation, cleaning, chunking, JSONL formatting, embeddings and vector storage form the current implementation work. Public supporting artefacts are still to be attached.

Retrieval & evaluation

Next: grounded retrieval, citations, missing-evidence behaviour and a representative evaluation set.

Controlled actions

Planned exploration: workflow orchestration and tool access with explicit boundaries and oversight.

05 / Evidence & limits

Be precise about what is proven.

This project page explains the current design and development direction. It does not yet include independently reviewable code artefacts, measured retrieval scores or production results. The interface above is illustrative.

Try the learning simulation

The linked lab uses a small synthetic corpus and deterministic rules. It demonstrates design choices; it is not the Caseworker implementation.

Open the retrieval lab

06 / Next steps

Test the questions that matter.

Build a representative set of answerable and unanswerable questions.
Inspect retrieved context and check whether it supports each answer.
Test permission boundaries and conflicting information.
Publish reviewed implementation artefacts and actual results.

Understand how retrieval works

A conversation starts with a real challenge

Explore the work.
Start a conversation.

Created by Absar Khan, connecting more than 20 years of technology experience with practical enterprise AI.

Discuss an opportunity About the approach

Caseworker.Evidence before answers.