Portfolio 16 · staged retrieval

RAG pipeline cost and capacity calculator

Derive chunk amplification exactly, separate initial build from recurring refresh and query serving, and let the slowest compatible measured stage bound sustainable throughput.

Decision this answers: Which stage drives monthly cost and which measured boundary throttles the pipeline before generation?

Scenario

Decision summary

Derived chunks

3529412exact-derived

Overlap amplification from exact stride.

[1]

Initial build

$1099.06exact-derived

Extraction, embeddings, and index writes.

[2][3]

Recurring monthly cost

$12927.92exact-derived

Refresh plus per-query serving stages.

[3]

Measured pipeline capacity

95.00measured

All stage measurements present: true.

[1]

Direct-labeled result profile

Derived chunks Initial build Recurring monthly cost Measured pipeline capacity

The tables below are the accessible source of truth; bar lengths never carry identity alone.

Pipeline stage ledger

Pipeline stage ledger
Candidate	Lifecycle	Units	Cost treatment	Capacity evidence	Boundary	Evidence
Extraction and chunking	Initial + refresh	Pages and derived chunks	Visible	Measured if timed	Exact positive stride	exact-derived[2]
Embedding and index write	Initial + refresh	Tokens and chunks	Visible	Measured if timed	ANN overhead delegated	exact-derived[3]
Retrieve, rerank, generate	Per query	Measured stage units	Visible	Slowest measured stage	Generation terminal only	measured[1][3]

Official surface versus this site

Retrieval-Augmented Generation paper

Use the official surface to confirm native prices, counting rules, limits, and product eligibility.

This offline workbench

The paper defines the retrieval-plus-generation architecture. This page adds a local operational ledger for extraction, chunking, embeddings, index writes, retrieval, reranking, retries, and terminal generation.

Sources and methodology

Every result-affecting reference is visible here without JavaScript and is retained in the JSON export.

[1] Retrieval-Augmented Generation paper

Authority: Meta AI
Native identifier: arXiv:2005.11401
Unit: retrieval plus generation stages
Evidence: exact-derived
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: The architecture informs stage boundaries; operational rates and capacity remain user measured.

Open the authoritative source

[2] Google Document AI pricing

Authority: Google Cloud
Native identifier: Document AI processing
Region: US
Unit: USD/pages
Evidence: published list
Price kind: list
As of: 2026-08-01
Effective from: 2026-08-01
Retrieved: 2026-08-01
Freshness: fresh
Normalization: The extraction rate is a dated editable scenario meter and remains separate from embeddings.

Open the authoritative source

[3] Vertex AI generative AI pricing

Authority: Google Cloud
Native identifier: embedding and generation meters
Region: global
Unit: USD/tokens
Evidence: published list
Price kind: list
As of: 2026-08-01
Effective from: 2026-08-01
Retrieved: 2026-08-01
Freshness: fresh
Normalization: Embedding and terminal generation meters are kept independent; no API-vs-hosting TCO is reproduced.

Open the authoritative source

Limits and non-claims

Raw vector bytes are visible conceptually, but ANN engine overhead belongs in the vector index sizer.
Every capacity result requires a compatible measured stage value; marketing throughput is not used.
Generation is a transparent terminal stage, not a duplicate of the inference TCO calculator.