Independent developer · Lebanon & remote GCC delivery
RAG knowledge assistants for approved business information.
A retrieval-augmented generation (RAG) assistant searches a defined collection of documents before preparing an answer. I scope and build document-based assistants for teams in Lebanon and the GCC, starting with one knowledge collection and the people allowed to use it.
Discuss your projectShare a short outline. I’ll review the fit, discuss your tools and requirements, then propose a paid scope before implementation.
See my personal RAG and AI-agent toolkit walkthrough →Start with the business problem.
Use retrieval when the answer depends on your documents rather than a general model response. Start by checking whether ordinary search is sufficient. For an assistant, agree which sources are authoritative, who maintains them and when the system should decline to answer.
What we can scope
- Internal policy and procedure lookup
- Support-agent assistance from approved product documentation
- Technical knowledge search across maintained documents
- A small pilot evaluated against representative staff questions
What the delivery can include
- Source inventory, document preparation and update requirements
- Retrieval design and an answer interface with source references
- Permission boundaries and tests for missing, stale or conflicting information
- Evaluation questions, an escalation route and a usage-cost discussion
Illustrative workflow · not a client case study
Example: an internal knowledge assistant
A team member asks about a procedure. The assistant searches information that person may access and drafts an answer with source references. This is an illustrative workflow, not a claim of a client deployment.
- Receive a staff question
- Check the user and permitted knowledge sources
- Retrieve relevant document passages
- Draft an answer with source references
- Decline or escalate when evidence is insufficient
What needs engineering: Test known-answer, no-answer, conflicting-source and restricted-document questions. Agree who maintains the collection and how updates or deleted documents reach the index.
If the assistant must update a system or carry out a task, that adds agent permissions, action validation and approval requirements. When retrieval also needs tool actions →
What shapes the project quote?
- Document formats, collection size, source quality and update frequency
- User access rules, source references and evaluation requirements
- Embedding and model usage, storage, hosting and ongoing maintenance
Scope and boundaries
The initial scope should name the collection, users and supported questions. Scanning or cleaning documents, bespoke connectors and model training are separate requirements if needed. Retrieval does not guarantee correct answers or replace permission checks.
Is this a useful first project?
Bring sample documents you have permission to use, typical questions and an owner who can judge answers. A small pilot should establish usefulness before expanding access. Sensitive data handling, hosting and provider terms need explicit agreement.
Relevant engineering work
The personal AI-agent toolkit contains document chunking, ChromaDB collections, source metadata and a query engine, alongside agent orchestration code. Its walkthrough distinguishes source inspection from runtime validation and describes checks needed before business use. This proof covers retrieval — finding and citing source-tagged context — not permissioned action execution; the toolkit's separate orchestrator is a distinct component, and neither is evidence that retrieved context has been wired to an approved tool action. See AI agent development for the orchestration/tool-permission side of that boundary.
Read the Personal RAG and AI-agent toolkit project walkthrough →Explore the project portfolio →Before we start
- How do documents stay current in a RAG assistant?
- Define an owner and an update process covering ingestion, re-indexing, deletion and source versions. Test whether changed and removed documents affect retrieval as intended. An answer should point to useful source material; a citation alone does not prove the answer is supported.
- Can a RAG assistant support Arabic and English documents?
- This can be scoped, but it needs representative documents and questions in both languages. Check extraction quality, mixed-language retrieval, terminology and answer usefulness with suitable reviewers. English-only tests do not establish Arabic or dialect performance.
- Where would our documents and AI requests be processed?
- That depends on the selected storage, embedding, model and hosting providers. Identify each processor, region, retention setting and access boundary before connecting sensitive information. A private interface does not by itself mean all processing stays on your premises or in your country.
- Does RAG guarantee correct answers?
- No. Retrieval can supply relevant context, but the assistant can still miss a document or produce an unsupported answer. Evaluate retrieval and answer quality separately, show useful source references and provide a route to a person.
- Is retrieval the same as training a model?
- No. Retrieval supplies selected information when a question is asked. It does not by itself train a model on your documents. Storage, provider processing and any training use must be checked for the chosen setup.
- Can different employees have different access?
- That should be a scoped requirement. Authorization must apply before information reaches the model, with tests for cross-user leakage. A metadata filter in a prototype is not proof of production access control.
Bring a workflow or application brief.
Share your company, current tools and the work you want to improve. Smaller projects are welcome. We can discuss fit, scope, integrations and a quote before work starts.
Describe your project