Fatbunny Digital
Discuss your project

Applied AI / retrieval engine

Give every agent a trustworthy way to find the information it is allowed to use.

Permission-aware retrieval, source authority, reranking and evaluation designed as shared business infrastructure.

We build retrieval engines that combine structured search, semantic retrieval, metadata, permissions and evidence checks so agents can use company information without treating every document as equally reliable.

Delivery
Permission-aware
Scope
Evidence-backed
Control
Reusable infrastructure

Retrieval operating path

Answer with evidence
  1. 01Question
  2. 02Permission
  3. 03Retrieve
  4. 04Rerank
  5. 05Check
  6. 06Answer

The operating problem

A model cannot know which company source is current, authoritative or permitted. Retrieval becomes useful only when search quality and information governance are designed together.

System mechanism

The evidence-backed knowledge path

A business question is answered only after identity, source authority, retrieval quality and evidence coverage agree.

A business question passes through permission checks, approved information retrieval, evidence evaluation and either a knowledge owner or an answer with sources.

  1. 01
    Business inputBusiness question

    A customer or employee asks a bounded question in a known operating context.

  2. 02
    Access policyPermission boundary

    Identity and role determine which sources may be searched.

  3. 03
    Retrieval corpusApproved information

    Current documents, records and product knowledge form the searchable corpus.

  4. 04
    Hybrid retrieval + rerankingFind the best evidence

    Hybrid search and reranking select relevant passages instead of returning the nearest text alone.

  5. 05
    Evidence gateEnough reliable evidence?

    Freshness, authority and coverage determine whether the system may answer.

  6. 06
    Human ownershipKnowledge owner

    Missing or conflicting evidence becomes an accountable content task.

  7. 07
    Business valueAnswer with sources

    The user receives an evidence-backed answer, clarification request or correct owner.

Fatbunny responsibility

Fatbunny defines the searchable corpus, permissions, metadata, retrieval strategy, reranking, answer boundary, source display, evaluation set and ownership process for missing knowledge.

Capabilities

  1. 01

    Corpus, metadata and source-authority design

  2. 02

    Hybrid keyword and semantic retrieval

  3. 03

    Permission-aware search and filtering

  4. 04

    Reranking, citations and retrieval evaluation

  5. 05

    Reusable retrieval APIs for agents and interfaces

Delivery decisions

  1. 01

    Choose source authority before chunking strategy

  2. 02

    Apply permissions before results reach a model

  3. 03

    Evaluate retrieval separately from answer quality

  4. 04

    Return insufficient evidence instead of generating around it

Expected outputs

  1. 01

    A governed searchable corpus and retrieval contract

  2. 02

    A working retrieval API with permission and source metadata

  3. 03

    Evaluation cases, content-gap reporting and operating guidance

Honest limits

  1. !

    Retrieval cannot repair inaccurate, contradictory or ownerless source information.

  2. !

    Vector similarity is not proof of authority or factual support.

  3. !

    Sensitive sources require identity, access, retention and audit controls outside the model.

Retrieval is shared agent infrastructure

LangGraph, Mastra, sales assistants and service systems all need a consistent way to find approved information. Building that capability once reduces duplicated search logic and governance gaps.

Search quality needs more than embeddings

Useful retrieval combines lexical and semantic signals, metadata, source authority, freshness and reranking. The best architecture depends on the corpus and the decisions the evidence must support.

Insufficient evidence is a valid result

A production system must be able to say that the available sources do not support an answer, request clarification or route the gap to the responsible content owner.

Direct answers

Common questions

What is the difference between RAG and a retrieval engine?

RAG describes using retrieved information during generation. A retrieval engine is the wider governed search capability: corpus, indexing, metadata, permissions, ranking, evidence and evaluation.

Do we need a vector database?

Not always. Many corpora benefit from hybrid search, and some are better served by an existing search engine or database with strong metadata and lexical relevance.

Can one retrieval engine support multiple agents?

Yes. A shared retrieval contract can support internal assistants, LangGraph or Mastra workflows, sales conversations and customer service while applying different permissions and source scopes.