Search results must respect the person asking
A knowledge assistant needs the same access boundaries as the documents behind it. Here is how to place those checks without making the model responsible for security.
Read articleCobnex blogs / Global
Cobnex blog
AI implementation, software architecture and cloud operations for teams worldwide.
146 articles in AI engineering
Page 1 of 9
When keyword search beats a vector match
Product codes and everyday questions need different retrieval behaviour. Hybrid search combines exact wording with meaning, but the combination needs a clear reason to exist.
Read articleA table cell is not a complete answer
Splitting a document into searchable passages can remove the information that makes a sentence or number meaningful. Choose boundaries around the questions the source needs to answer.
Read articleA citation can be real and still support the wrong claim
Linking to a genuine document does not prove an answer is correct. The evidence must support the particular claim, including its conditions and exceptions.
Read articleDeleting a policy should also retire its cached answers
Removing a document from search does not remove the answers already built from it. Design freshness around the complete route from source to reader.
Read articleSupplier documents must not become agent instructions
An assistant may need to read an external document without accepting its requests. Keep the authority to act in application code and authenticated user intent.
Read articleGive an agent a small tool, not an administrator console
A useful agent capability describes a business operation with clear limits. General administrative access makes mistakes harder to contain and results harder to explain.
Read articleAn approval should authorise one exact proposal
A reviewer needs to approve the action that will actually run. Tie the decision to a saved target, payload and version so later changes cannot inherit it.
Read articleResuming an agent is more than replaying its conversation
Conversation history explains intent. Durable operation records establish what happened. Recovery needs both, with a clear rule for uncertain side effects.
Read articleA better demo is not an evaluation dataset
A handful of polished examples can show what an assistant might do. A useful evaluation set shows where it works, where it fails and what those failures mean.
Read articleRelease a model change like a behaviour change
A replacement model can keep the same API shape while changing answers, tool choices and refusals. Assess those behaviours as part of the application release.
Read articlePrivate inference has more than one data boundary
A private network connection addresses one part of an AI system. Prompts, retrieval stores, diagnostics and support access still need their own data-handling decisions.
Read articlePrice the completed AI task, including its retries
A task budget should cover the whole attempt to produce a useful result. Model calls, tools, validation and recovery all draw from that budget.
Read articleKnowing when an assistant should decline to answer
Missing evidence, unclear intent and missing authority need different responses. Give the product a useful path for each instead of one generic refusal.
Read articleCarry document permissions into the search index
Indexing protected documents is a data-modelling job as well as a search job. Build the ingestion path so content and access rules stay attached to each other.
Read articleBuild the lexical baseline before adding embeddings
Start with a search path you can explain. A working keyword baseline makes it easier to see what semantic retrieval adds and where it introduces mistakes.
Read articlePreserve heading ancestry when splitting reports
A passage needs to retain the section that gives it meaning. Build a small document model before producing the records sent to the search index.
Read articleCarry source identifiers through answer generation
Build citations from records owned by the application. Keep a generated answer from deciding which URLs or document versions count as its evidence.
Read article