Skip to content
Emmanuel NanadoumRésumé(PDF, opens in a new tab)

KnowledgeOS

Ask the company. See the evidence behind the answer.

Explore enterprise product Try KnowledgeOS Request enterprise demo
Built for
Portfolio implementation — enterprise knowledge assistant
Engagement
Retrieval-augmented generation / AI implementation
My role
Requirements, architecture, retrieval design, grounding, API implementation, evaluation design and deployment
Source
emmanuel-nanadoum.vercel.app/knowledgeos
KnowledgeOS architecture: documents to embeddings, retrieval, grounded generation, and cited answers.
emmanuel-nanadoum.vercel.app/knowledgeos

Status at time of writing

Stamped from the live build, not from the plan.

Status of each component
ComponentStatus
Interactive RAG demoLive

Question, retrieval, grounded answer and source evidence are implemented.

Embedding retrievalLive

Gemini embeddings when configured, with deterministic lexical fallback.

Enterprise multi-tenant controlsNot activated

Production architecture is documented; organization auth, RLS and ingestion are the next product phase.

1Business problem

Company knowledge is fragmented across policies, playbooks, implementation guides and operational systems, forcing employees to search manually or rely on memory.

A generic chatbot can answer confidently without evidence. For business policy and operations, an unsupported answer can be worse than no answer.

2Discovery

  1. 2.1The useful unit is not merely an answer; it is an answer plus the evidence that supports it.
  2. 2.2Unsupported policy generation is a failure mode, so refusal behavior is a product requirement rather than an edge case.
  3. 2.3A synthetic demonstration corpus proves the architecture without exposing private company data.

3Requirements

Requirements
R-01Retrieve relevant evidence before generation.
R-02Use semantic embeddings when available with a resilient deterministic retrieval fallback.
R-03Constrain generation to retrieved evidence and refuse unsupported questions.
R-04Keep model credentials and AI execution server-side.
R-05Expose sources, architecture, limitations and production extension points.
R-06Design the production path for precomputed vectors, Postgres/pgvector and permission-aware retrieval.

4Solution architecture

KnowledgeOS separates retrieval from generation so the model receives a bounded evidence set before it is allowed to answer.

Live RAG path

  1. 01User question
  2. 02Query embedding
  3. 03Retrieve + rank
  4. 04Grounded generation
  5. 05Answer + citations
Figure 4.1 — Live RAG path

Production ingestion path

  1. 01Upload
  2. 02Parse + normalize
  3. 03Chunk + metadata
  4. 04Embeddings
  5. 05Postgres / pgvector
  6. 06Permission filter
  7. 07Evaluation + monitoring
Figure 4.2 — Production ingestion path

5Demo / POC

The public demo includes answerable questions and an intentionally unsupported parental-leave question to make refusal behavior visible.

KnowledgeOS retrieval-augmented generation architecture.
Figure 5.1 — Evidence is retrieved before generation; unsupported answers are refused.

6Technologies

  • Next.js
  • TypeScript
  • Gemini API
  • Gemini embeddings
  • RAG
  • REST API
  • Vercel Functions
  • Source citations
  • Lexical fallback
  • Postgres / pgvector production design

7Integration points

Integration points
FromTo
Browser/api/knowledgeosQuestion in; grounded answer, retrieval mode and sources out.
RAG APIEmbedding modelSemantic query and knowledge representations when configured.
Retrieved evidenceGenerative modelOnly ranked source context is supplied for answering.
Production designPostgres / pgvector + RLSPrecomputed vectors and permission-aware retrieval.

8Tradeoffs

Design decisions and tradeoffs
Synthetic public corpusInstead of: Private company documentsDemonstrates the system safely without exposing customer or employer information.
Evidence-first refusalInstead of: Always produce an answerTrust requires the system to expose when the knowledge base cannot support a claim.
Embeddings plus lexical fallbackInstead of: Single-provider dependencyThe demo remains testable if embedding service availability changes.
Compact live corpusInstead of: Premature ingestion platformProves the core RAG loop first; production ingestion is the next product phase.

9Implementation & handoff

  • Production rollout adds authenticated organization workspaces, document upload, asynchronous parsing/chunking and precomputed embeddings.
  • Postgres/pgvector plus row-level security provides organization and user-aware retrieval boundaries.
  • Evaluation sets, latency/cost monitoring, source administration and controlled re-indexing become operational controls.

10What this demonstrates

  1. 1.Implements the complete retrieval → ranking → grounded generation → citation/refusal loop.
  2. 2.Treats hallucination control, evidence visibility and failure behavior as implementation requirements.
  3. 3.Documents production architecture and tradeoffs without presenting unbuilt enterprise controls as shipped.