Kunnworks

Find knowledge. Carry out approved work.

Answer questions with evidence from your work materials, then carry out actions after a person approves them

Answers grounded in work documentsIllustrative scenario

This request is for 12 boxes of gloves. Can it be dispatched?

Source-based answer

Check approval and delivery address as well as the item and quantity. The available material has no approval record. Ask the dispatch approver to review it.

Referenced work document
Dispatch manual §2 — Verify item, quantity, delivery address and approval before dispatch.
Next human actionRequest approval review
Illustrative service workflowFeature illustration
IMPLEMENTATION SCOPE

Permission-aware AI from search to action

Build retrieval and generation that preserve document permissions and versions. Define tool validation, allowed actions, approval, stopping and retry conditions separately.

What we build
Data connectors · retrieval/agent flows · tool APIs · evaluation set
What we validate
Missing evidence, revoked access, malicious document instructions, duplicate actions and expired approvals
Read the architecture, acceptance scenarios and handover details
IN YOUR EVERYDAY WORK

From finding an answer
to carrying out approved work.

Before listing features, we define the inputs, decisions and next actions in a real working situation.

An illustrative scenario. Specific processing rules are agreed for the project.

WORKFLOW EXAMPLEILLUSTRATIVE

An employee checks return procedures and requests a task.

Input
Permitted manual · customer request
Review
Answer evidence and action permissions
NEXT ACTION

Review the evidence and execute only the approved action

When this service helps

When people repeatedly search the same material or repeat actions across systems

  1. 1

    An employee asks a work question

  2. 2

    Check permitted sources or actions

  3. 3

    Review the answer or approve execution

What the project delivers

  • Search and question-answering interface
  • Permission-aware source integration
  • Approval flow and execution records

Specific features, integrations and operating support are agreed in the project proposal

Technical design and acceptance criteria

Details for technical reviewers. Architecture and acceptance criteria are agreed for the actual environment and project scope.

  • Hybrid retrieval
  • Reranking
  • Tool contracts
  • Task-level evaluation

Processing flow and exceptions

Use the query and caller permissions to retrieve lexical and vector candidates, fuse rankings, and rerank context. Separate an answer from an action: tools that change external state pass through explicit policy and approval.

Authorized retrieval and generationUse the query and caller permissions to retrieve lexical and vector candidates, fuse rankings, and rerank context. Separate an answer from an action: tools that change external state pass through explicit policy and approval.Authorized retrieval and generationKUNNWORKS · DESIGN EXAMPLERetrieval with document ACLsQueryBM25VectorRRFRerankLLMEvaluate retrievalBounded contextAnswerCitationsTool actions pass separate policy and approval checks; documents never grant permissions
AI AGENTS / KNOWLEDGE TO ACTIONDesign example · Refined against scope and constraints

Architecture summary

  1. Retrieval quality — ACL filters, hybrid candidates, reranker
  2. Tool execution — Schemas, allowlists, approval, idempotency
  3. Evaluation and operation — Question set, grounding, completion, execution budgets
02Acceptance criteria

Implementation and acceptance criteria

Concrete design decisions and acceptance scenarios to agree for the project.

Retrieval quality

ImplementationACL filters, hybrid candidates, reranker

ValidationInternal terms, tables, revisions, revoked access

Tool execution

ImplementationSchemas, allowlists, approval, idempotency

ValidationReplay, hostile document instructions, expired approval

Evaluation and operation

ImplementationQuestion set, grounding, completion, execution budgets

ValidationReproduce retrieval and generation failures separately

Technology choices and acceptance criteria depend on discovery, integrations, and agreed scope. We define reproducible acceptance scenarios and handover evidence for the operating environment.

03Implementation details

Engineering decisions
in detail

From visible features to failure conditions,
we work through the decisions before launch.

01

Preserve source permissions and versions through ingestion

Retain source ID, location, revision, section, and ACL in chunk metadata. Respect tables, lists, and headings when chunking, and propagate deletions and permission changes through incremental indexing. Enforce retrieval filters and recheck source access; the model must never decide whether the caller is allowed to see a document.

02

Evaluate retrieval separately from generation

Combine lexical and embedding candidates using a method such as reciprocal rank fusion, then assess whether cross-encoder reranking improves the task. Test Korean, English, and internal terminology using retrieval coverage and rank quality. Separately evaluate grounding, citation validity, abstention, and task completion, so a plausible answer cannot hide a failed search.

03

Tool calls need schemas, idempotency, and approval

Validate tool inputs against typed schemas and restrict API and data scope. For side effects such as posting a record, sending mail, or changing access, define idempotency keys, approvers, and execution expiry. Treat prompt injection through layered controls: untrusted-content separation, least privilege, and pre-execution policy checks, not a single text filter.

04

Operate through a traceable execution graph

Correlate retrieval candidates, chosen context, prompt and model versions, tool calls, and approval results. Apply redaction and retention rules to sensitive traces while separating latency, cost, retries, and failures. Set step, token, and tool budgets with cancellation and human-handoff criteria to bound loops and uncontrolled execution.

04Handover

Detailed deliverables

  • Document connectors, access and retrieval design, and an evaluation set
  • Agreed agent flows, tool APIs, and approval steps
  • Evaluation results, operating metrics, and escalation guidance
05Before implementation

Questions before implementation

  • The target workflow and what the agent may actually change
  • Document locations, formats, permissions, and refresh frequency
  • Human approval rules and success or stop criteria

How your project moves forward

01

Discover

Understand your goals, workflows, and operating environment.

02

Define

Define features, priorities, and integration requirements.

03

Build

Build the agreed scope and review progress together.

04

Launch & Operate

Prepare testing, deployment, and operational handover.

Let’s discuss your project

Tell us what you want to build or improve
Our team will review your request and reply by email

Discuss your project