RETRIEVAL AUGMENTED GENERATION / APPLICATION AILive product

Huck

Your information. Answers you can inspect.

huck.chat

Huck’s live homepage with its illustrated frontier character and the headline “Your data. AI answers. With receipts.” — The live Huck homepage. The character is the product’s identity; customers can choose their own assistant name and appearance.

Huck is a BWD product, designed and built by Matthew Newport. It connects an assistant to the information a business actually runs on. Applications send structured records as they change; each answer uses the context the visitor is allowed to see. The product brings data synchronization, access rules, conversation and release testing into one system.

The problem

A convincing answer is not enough. Business records change, people have different access, and a new prompt can quietly break a previously useful answer. The assistant needs current evidence, server-enforced permissions and a repeatable way to test a change before it reaches visitors.

The implementation

A Laravel API accepts typed records with nested JSON content, revision checks and optional access gates. PostgreSQL and pgvector support retrieval; the application checks authorization before sending evidence to the model. Filament panels bring knowledge, access hierarchies, prompt versions and evaluation results together. Signed action webhooks let a customer’s system report progress and results back into a conversation.

API synchronization for structured records, nested content and typed relationships
Customizable access hierarchies with server-enforced retrieval boundaries
Versioned prompts, golden cases and deterministic or model-based judges using authorized context
Cited answers, insufficient-evidence responses and configurable website topic boundaries
An embeddable assistant with configurable name, character, appearance and placement
Signed action webhooks with progress, completion and conversation continuation

01Sync current records02Retrieve within permissions03Generate and cite an answer04Evaluate before release

02 / ENGINEERING DECISIONS

Where the details matter.

Access is an application rule

The model does not decide who may see a record. Credentials identify the workspace; trusted subject relationships and configured gates constrain retrieval. A record’s relationship to another record does not grant access to it.

Test the exact candidate

A release needs a complete passing evaluation of its current prompt, authorized context, policy and required golden suite. A source or policy change invalidates old evidence. Expected answers are judge inputs, never hints supplied to the answering model.

A character is a choice

Huck gives the product a recognizable identity. The embedded assistant can take the customer’s own name and presentation, while configurable topic boundaries and input/output checks keep its role focused on that business.

03 / PACKAGES & CODE

Inside the implementation.

The application owns permissions, website scope and release approval. These source excerpts show checks at the retrieval and publication boundaries.

APPLICATION EXCERPT

Check the whole request before retrieval

A scoped assistant checks the question and actual conversation history before retrieving business records. A negative decision returns the configured refusal without searching for evidence.

if ($scope !== null) {
    $history = $scopeService->history($history);
    $scopeChecks[] = $scopeService->check($w, $purpose, $scope, 'input', ['question' => $question, 'prior_conversation' => $history], $scopeGuard);
    if (! $scopeChecks[0]['allowed']) {
        return $this->scopeRefusal($prompt, $epoch, $sv, $scopeChecks);
    }
}

Actual application excerpt · 38758fa · lines 59–65. Provider failures fail closed in the scope service; retrieval follows this block.

APPLICATION EXCERPT

Bind approval to current evaluation evidence

Publication compares the current manifest with the completed run and requires the exact golden case identities. Every case must have completed and passed. Earlier checks also verify the candidate, retained context and current-context mode.

$requiredIds = array_column($manifest['goldens'], 'id');
$resultIds = array_column($run->results, 'golden_id');
sort($requiredIds);
sort($resultIds);
abort_unless($requiredIds === $resultIds, 409, 'Evaluation case identities do not match the required suite.');
foreach ($run->results as $result) {
    abort_unless(($result['passed'] ?? false) === true && ($result['status'] ?? '') === 'completed', 409, 'Every required case must pass.');
}

Actual application excerpt · 38758fa · lines 149–156. The surrounding transaction locks the workspace, prompt and evaluation run.

04 / VERIFICATION & CURRENT SCOPE

What holds it together.

Authorization tests cover cross-workspace access, incomplete hierarchy identities and access revocation during a request.
Evaluation and release tests reject stale evidence, incomplete suites, expired context and individual failed cases.
Website-scope tests cover mixed requests, unrelated questions, prompt-disclosure attempts and provider failures that must fail closed.
The screenshot is a capture of the live product, not an illustration of an unbuilt interface.
The public API and developer guides document structured knowledge, keys, sessions, integration examples and failure handling.

CURRENT RELEASE

Huck is live at huck.chat. Visit Huck for current plans and subscription availability. Evaluations and citations help assess answer quality; they do not guarantee that every AI response is correct.

KEEP EXPLORING