Project 1 — Production RAG SaaS

Project 1 — the production RAG SaaS — is the multi-tenant knowledge platform, built from the schema to the evaluation (L381) — the scope (the requirements L359: the tenants L320, the documents L176, the queries L358, L381), the architecture (the L349 platform L349: the ingestion L176, the retrieval L189, the isolation L320, L381), the data model (the schema L381: the tenants L320, the documents L265, the chunks L178, the vectors L183, L381), and the evaluation (the evals L341: the retrieval L338, the groundedness L337, in the CI L296, L381) — the RAG's (L349) proof (L381).

The one-sentence interview answer: "Project 1 is the RAG SaaS, built end to end (L381). The scope (L381): the requirements (L359) — the tenants (L320): the multi-tenant (L357) knowledge (L381); the documents (L176): the PDFs (L177) and the markdown (L381); and the queries (L358): the natural language (L348) over the knowledge (L381). The architecture (L381): the L349 platform (L349) — the ingestion (L176): the S3 (L265) → the queue (L270) → the parse (L177), the chunk (L178), the embed (L181); the retrieval (L189): the pgvector (L183) with the hybrid (L187) and the rerank (L190); and the isolation (L320): the per-tenant (L320) indexes (L183) and the filters (L189). The data model (L381): the schema (L381) — the tenants (L320), the documents (L265), the chunks (L178), and the vectors (L183). The build (L381): the pipeline (L349), the API (L267), and the frontend (L96). The evaluation (L341): the evals (L341) — the retrieval (L338): the precision and the recall (L338) on the golden set (L342); and the groundedness (L337) — in the CI (L296), gating (L341). The RAG's (L349) proof (L381) — from the schema (L381) to the evaluation (L341)."

Think of the RAG SaaS as the knowledge library, built. The builder (the architect, L381) builds the library (L381): the blueprint (the scope, L381) — the branches (the tenants, L320) and the collections (the documents, L176); the intake (the ingestion, L176) — the donations (the PDFs, L177) sorted (L177), cut (L178), and catalogued (L181); the shelves (the index, L183) — the per-branch (L320); the desks (the retrieval, L189); and the inspectors (the evals, L341) — the reference checks (L338). The library works because the blueprint is clear, the intake runs, and the inspectors gate (L381).

   the knowledge library (the RAG SaaS, L381)
   ┌────────────────────────────────────────────────────────┐
   │ the blueprint (the scope, L381) · the intake (the      │
   │ ingestion, L176) · the shelves (the index, L183)       │
   │ the desks (the retrieval, L189) · the inspectors (the  │
   │ evals, L341)                                           │
   └────────────────────────────────────────────────────────┘

The mental model is the library: the blueprint, the intake, and the inspectors (L381).

   THE SCOPE (L381) → the requirements (L359)
        │
        ▼
   THE DATA MODEL (L381) → the schema (L381)
        │
        ▼
   THE ARCHITECTURE (L381) → the L349 platform (L349)
   the ingestion (L176) → the retrieval (L189) → the isolation (L320)
        │
        ▼
   THE EVALUATION (L341) → the evals (L338, L337) in the CI (L296)

The flow is the build: scope → schema → architecture → evaluation (L381).

• The scope (L381). The requirements (L359): the tenants (L320), the documents (L176), the queries (L358).
• The architecture (L381). The L349 platform (L349): the ingestion (L176), the retrieval (L189), the isolation (L320).
• The data model (L381). The schema (L381): the tenants (L320), the documents (L265), the chunks (L178), the vectors (L183).
• The evaluation (L341). The evals (L341): the retrieval (L338), the groundedness (L337), in the CI (L296).

• The portfolio (L103). Project 1 (L381) — the L349 proof (L381).
• An interview (L381). The walkthrough (L381) — the schema to the eval (L381).
• A knowledge product (L357). The RAG SaaS (L381) — the tenants (L320) and the knowledge (L349).
• A support copilot (L350). The RAG (L381) — the help center (L265) grounded (L337).
• Anything RAG (L349). The capstone (L381) — the L349 platform (L349), built (L381).

The through-line: the proof is the RAG's — the L349 platform (L349), from the schema (L381) to the eval (L341).

Say it in four moves:

1. The scope. "The tenants (L320), the documents (L176), the queries (L358)."
2. The architecture. "The L349 platform (L349) — the ingestion (L176), the retrieval (L189), the isolation (L320)."
3. The data model. "The schema (L381) — the tenants, the documents, the chunks, the vectors (L183)."
4. The evaluation. "The evals (L341) — the retrieval (L338) and the groundedness (L337) in the CI (L296)."

• The schema is the foundation (L381). The data model (L381) — the tenants (L320) and the chunks (L178) — the L370 seam (L370), schema-shaped (L381).
• The isolation is the SaaS's (L320). The per-tenant (L320) indexes (L183) and the filters (L189) — the L357 wall (L357), built (L381).
• The evals are the gate (L341). The retrieval (L338) and the groundedness (L337) — the L341 suite (L341) in the CI (L296) — the quality (L341) gated (L381).
• The pipeline is the async (L176). The queue (L270) — the ingestion (L176) — the L222 engine room (L222), built (L381).
• The observability is the standard (L346). The OTel (L346) — the tokens (L332) and the cost (L334) — the L346 standard (L346), in the capstone (L381).

• The demo RAG (L381). The prompt-only (L349) — the pipeline (L176) and the schema (L381) missing (L381).
• The schema-less (L381). The vectors (L183) without the tenants (L320) and the chunks (L178) — the data model (L381) first (L381).
• The shared index (L320). The one index (L183) — the cross-tenant (L320) — the isolation (L320) (L381).
• The eval-less (L341). The retrieval (L338) and the groundedness (L337) un-gated (L296) — the quality (L341) unknown (L381).
• The un-observed (L346). The tokens (L332) and the cost (L334) un-measured (L346) — the L346 standard (L346) (L381).

• Design the schema first (L381) — the tenants (L320), the chunks (L178), the vectors (L183).
• Build the pipeline (L349) — the ingestion (L176) and the retrieval (L189).
• Isolate the tenants (L320) — the per-tenant (L320) indexes (L183).
• Gate the evals (L341) — the retrieval (L338) and the groundedness (L337) in the CI (L296).
• Observe the platform (L346) — the OTel (L346) with the cost (L334).

Q: Walk me through Project 1.

A: The RAG SaaS, built end to end (L381). The scope — the tenants (L320), the documents (L176), the queries (L358). The architecture — the L349 platform (L349): the ingestion (L176), the retrieval (L189), the isolation (L320). The data model — the schema (L381). And the evaluation — the evals (L341) in the CI (L296).

Q: What's the data model?

A: The schema (L381): the tenants (L320) — the multi-tenant (L357) root; the documents (L265) — the S3 (L265) sources; the chunks (L178) — the parsed (L177) and the chunked (L178); and the vectors (L183) — the pgvector (L183) embeddings (L181). The schema (L381) is the L370 seam (L370) — the per-tenant (L320) from the start (L381).

Q: How do you evaluate it?

A: The L341 suite (L341): the golden set (L342) — the queries (L338) with the relevant docs (L342); the retrieval (L338) — the precision and the recall (L338); and the groundedness (L337) — the L337 check (L337). The suite (L341) in the CI (L296) — gating the deploy (L307) (L381).

Q: How do you isolate the tenants?

A: The L320 wall (L320): the tenant ID (L320) in the indexes (L183) and the filters (L189) — the per-tenant (L320) retrieval (L189); the per-tenant (L320) keys (L269) in the cache (L269); and the per-tenant (L320) quotas (L149) in the gateway (L267) — the L357 wall (L357), built (L381).

• What's the scope (L381)?
• What's the data model (L381)?
• How do you evaluate it (L341)?
• How do you isolate the tenants (L320)?
• What's the arc (L381)?

The senior read: the right column is the capstone — the schema to the eval (L381).

// Project 1 (L381) — the RAG SaaS, from the schema (L381).
// 1 · THE SCHEMA (L381) — the data model (L381).
const schema = {
  tenants:  { id: 'uuid', name: 'text' },              // L320
  documents: { id: 'uuid', tenantId: 'fk', s3Key: 'text' },   // L265, L320
  chunks:   { id: 'uuid', documentId: 'fk', index: 'int', text: 'text' },  // L178
  vectors:  { chunkId: 'fk', embedding: 'vector(1536)' },  // L181, L183
};

// 2 · THE PIPELINE (L349) — the ingestion (L176).
async function onDocumentUploaded(event) {             // the S3 event (L276)
  await queue.enqueue(event.documentId);               // the queue (L270)
}
async function worker(documentId) {                    // L266
  const text = await parse(documentId);                // L177
  const chunks = chunk(text);                          // L178
  const vectors = await embed(chunks);                 // L181
  await index.write(chunks, vectors, { tenantId });    // L183, L320
}

// 3 · THE RETRIEVAL (L189) — with the isolation (L320).
const top = await index.search(tenantId, query, { topK: 20 });   // L189, L320
const reranked = await rerank(top, query);             // L190

// 4 · THE EVALS (L341) — in the CI (L296).
//   the golden set (L342): the retrieval (L338) >= 0.85 recall (L381)
//   the groundedness (L337) >= 0.9 (L381)

What the reader must SEE — the build, started:

  the tenants + chunks + vectors schema → the data model (L381)
  the queue + the parse + the embed     → the pipeline (L176)
  the per-tenant search + the rerank    → the retrieval (L189)
  the golden set in the CI              → the evals (L341)

  From the schema to the evaluation (L381).

• The pipeline is the async (L176). The queue (L270) — the ingestion (L176) — the L222 engine room (L222) (L381).
• The retrieval is the TTFT (L189). The index (L183) and the rerank (L190) — the sub-second (L151) (L381).
• The evals are the CI's (L341). The golden set (L342) — the minutes (L381) in the pipeline (L296).
• The cost is the tokens' (L332). The embedding (L181) and the generation (L278) — the L334 attribution (L334) (L381).

• Project 1 = the RAG's proof (L381): the scope, the architecture, the data model, the evaluation.
• The scope: the tenants (L320), the documents (L176), the queries (L358).
• The architecture: the L349 platform (L349) — the ingestion (L176), the retrieval (L189), the isolation (L320).
• The data model: the schema (L381) — the tenants, the documents, the chunks, the vectors (L183).
• The evaluation: the evals (L341) — the retrieval (L338) and the groundedness (L337) in the CI (L296).

PROJECT 1: PRODUCTION RAG SAAS = the multi-tenant knowledge platform

THE SCOPE (L381)
  the tenants (L320) — the multi-tenant (L357) knowledge (L381)
  the documents (L176) — the PDFs (L177), the markdown (L381)
  the queries (L358) — the natural language (L348)

THE ARCHITECTURE (L381) — THE L349 PLATFORM (L349)
  the ingestion (L176): the S3 (L265) → the queue (L270)
    → the parse (L177), the chunk (L178), the embed (L181)
  the retrieval (L189): the pgvector (L183), the hybrid (L187),
    the rerank (L190)
  the isolation (L320): the per-tenant (L320) indexes (L183)

THE DATA MODEL (L381)
  the tenants (L320) · the documents (L265) · the chunks (L178)
  the vectors (L183) — the schema (L381)

THE EVALUATION (L341)
  the retrieval (L338) — the precision and the recall (L338)
  the groundedness (L337) · the golden set (L342)
  in the CI (L296) — gating (L341)

INTERVIEW, 4 MOVES
  1 scope   "the tenants, the documents, the queries (L381)"
  2 architecture "the L349 platform (L349)"
  3 schema  "the data model (L381)"
  4 evals   "the golden set in the CI (L341)"

You now hold the first proof: the scope, the architecture, the schema, and the evaluation — with the intake running and the inspectors gating. The knowledge library is built — and it fills the portfolio (L381).

Next: the guarded, audited, approval-gated agent — Project 2 (L382).