Web app & API
$20/mo
Vercel · Pro
Next.js app and API routes with preview deploys and no servers to run
vercel.com ↗PDF Q&A with RAG
Users upload PDFs and ask questions; answers are grounded in the documents with page citations, using retrieval-augmented generation.
Good fit: Internal knowledge bases, contract or policy assistants, research tools: up to a few thousand documents per workspace.
- Infrastructure
- $258/mo
- $173 at half the usage · $780 at double
How the parts connect
Hover or click a part to see what flows between them. Drag to move, use the buttons to zoom.
- Web app & API$20/mo
Vercel · Pro
- Ingestion jobs$0
Inngest · Free
← ingest document from Web app & API
- AI framework$0
Vercel AI SDK
← built with from Web app & API
- PDF parsing$50/mo
LlamaParse · Starter
← PDF → text + pages from Ingestion jobs
- Embedding model$0.05/mo
OpenAI API · text-embedding-3-small
← chunks from Ingestion jobs · question from Web app & API
- Vector database$20/mo
Pinecone · Builder
← upsert vectors from Ingestion jobs · top-k search from Web app & API
- App database$25/mo
Supabase · Pro
← users, docs, chats from Web app & API
- File storage$0
Cloudflare R2
← uploaded PDFs from Web app & API
- Answer model$114/mo
Claude API · Claude Sonnet 5.5
← question + passages from Web app & API
- LLM observability$29/mo
Langfuse · Core
← traces from Web app & API
Monthly cost
EstimateExpected cost at the usage below, and the range from half to double that usage. Worked out from each vendor's current prices and free tiers; your real bill depends on your traffic.
| Part | Pick | Expected / mo | Range |
|---|---|---|---|
| Web app & API | Vercel · Pro | $20/mo | fixed |
| AI framework | Vercel AI SDK | $0 | fixed |
| PDF parsing | LlamaParse · Starter | $50/mo | $50–$225 |
| Embedding model | OpenAI API · text-embedding-3-small | $0.05/mo | $0.01–$0.20 |
| Vector database | Pinecone · Builder | $20/mo | fixed |
| App database | Supabase · Pro | $25/mo | fixed |
| File storage | Cloudflare R2 | $0 | $0–$0.72 |
| Ingestion jobs | Inngest · Free | $0 | fixed |
| Answer model | Claude API · Claude Sonnet 5.5 | $114/mo | $28.60–$458 |
| LLM observability | Langfuse · Core | $29/mo | $29–$31.24 |
| Infrastructure total | $258/mo | $173–$780 | |
Show the arithmetic behind every number
Web app & API · Vercel · Pro
- vcpu = app_vcpu = 0.5
- ram_gb = app_ram_gb = 0.5
- Pro list price $20.00/month
- note: includes $20 of usage credit; function, bandwidth and build usage beyond it is not modeled
AI framework · Vercel AI SDK
- free open-source software (Apache-2.0); runs inside the app
PDF parsing · LlamaParse · Starter
- pages = documents_per_month * pages_per_document = 300 × 15 = 4,500
- credits = documents_per_month * pages_per_document * parse_credits_per_page = 300 × 15 × 3 = 13,500
- Starter list price $50.00/month
- credits: 13,500 − 40,000 free = 0 ÷ 1,000 × $1.25 = $0.00
Embedding model · OpenAI API · text-embedding-3-small
- input_tokens = documents_per_month * pages_per_document * tokens_per_page + monthly_active_users * questions_per_user_per_month * question_tokens = 300 × 15 × 500 + 200 × 40 × 30 = 2,490,000
- input: 2,490,000 tokens ÷ 1M × $0.02 = $0.05
Vector database · Pinecone · Builder
- vector_gb = retention_months * documents_per_month * pages_per_document * tokens_per_page / chunk_tokens * vector_kb_per_chunk / 1000000 = 12 × 300 × 15 × 500 / 800 × 8 / 1000000 = 0.27
- db_gb = retention_months * documents_per_month * pages_per_document * tokens_per_page / chunk_tokens * vector_kb_per_chunk / 1000000 = 12 × 300 × 15 × 500 / 800 × 8 / 1000000 = 0.27
- write_units = documents_per_month * pages_per_document * tokens_per_page / chunk_tokens * vector_kb_per_chunk = 300 × 15 × 500 / 800 × 8 = 22,500
- read_units = monthly_active_users * questions_per_user_per_month * read_units_per_query = 200 × 40 × 1 = 8,000
- Builder list price $20.00/month
App database · Supabase · Pro
- db_gb = monthly_active_users * questions_per_user_per_month * retention_months * kb_per_chat_turn / 1000000 = 200 × 40 × 12 × 4 / 1000000 = 0.384
- mau = monthly_active_users = 200
- Pro list price $25.00/month
- database GB: 0.384 − 8 free = 0 × $0.125 = $0.00
- file storage GB: 0 − 100 free = 0 × $0.0213 = $0.00
- monthly active users: 200 − 100,000 free = 0 × $0.00325 = $0.00
File storage · Cloudflare R2
- storage_gb = documents_per_month * retention_months * mb_per_document / 1000 = 300 × 12 × 2 / 1000 = 7.2
- write_ops = documents_per_month = 300
- read_ops = documents_per_month + monthly_active_users * questions_per_user_per_month * citations_opened_per_question = 300 + 200 × 40 × 0.3 = 2,700
- storage GB-month: 7.2 − 10 free = 0 × $0.015 (catalog: Storage) = $0.00
- Class A (write) ops: 300 − 1,000,000 free = 0 ÷ 1,000,000 → 0 (rounded up) × $4.50 (catalog: Class A Operations) = $0.00
- Class B (read) ops: 2,700 − 10,000,000 free = 0 ÷ 1,000,000 → 0 (rounded up) × $0.36 (catalog: Class B Operations) = $0.00
Ingestion jobs · Inngest · Free
- executions = documents_per_month * (ingest_steps + 1) = 300 × (4 + 1) = 1,500
- job_runs = documents_per_month = 300
- job_compute_seconds = documents_per_month * pages_per_document * seconds_per_page = 300 × 15 × 2 = 9,000
- Free list price $0.00/month
Answer model · Claude API · Claude Sonnet 5.5
- input_tokens = monthly_active_users * questions_per_user_per_month * (top_k * chunk_tokens + prompt_overhead_tokens) = 200 × 40 × (6 × 800 + 600) = 43,200,000
- output_tokens = monthly_active_users * questions_per_user_per_month * answer_tokens = 200 × 40 × 350 = 2,800,000
- input: 43,200,000 tokens ÷ 1M × $2.00 = $86.40
- output: 2,800,000 tokens ÷ 1M × $10.00 = $28.00
LLM observability · Langfuse · Core
- units = monthly_active_users * questions_per_user_per_month * observations_per_question = 200 × 40 × 4 = 32,000
- traces = monthly_active_users * questions_per_user_per_month = 200 × 40 = 8,000
- seats = team_seats = 2
- Core list price $29.00/month
- units (traces + observations + scores): 32,000 − 100,000 free = 0 ÷ 100,000 × $8.00 = $0.00
The tools, and why
AI framework
$0
Vercel AI SDK
Streaming chat UI hooks and one API for every model provider, in TypeScript
ai-sdk.dev ↗PDF parsing
$50/mo
LlamaParse · Starter
Best table and layout extraction for messy, scanned or multi-column PDFs
llamaindex.ai ↗Embedding model
$0.05/mo
OpenAI API
Model: OpenAI text-embedding-3-small
Strong retrieval quality at the lowest per-token price in the catalog
platform.openai.com ↗Vector database
$20/mo
Pinecone · Builder
Serverless index with metadata filtering per workspace; no capacity planning
pinecone.io ↗App database
$25/mo
Supabase · Pro
Postgres plus auth and row-level security for per-workspace access
supabase.com ↗File storage
$0
Cloudflare R2
S3-compatible with no egress fees when users open their PDFs
cloudflare.com ↗Ingestion jobs
$0
Inngest · Free
Durable step functions: a failed embed retries without re-parsing the PDF
inngest.com ↗Answer model
$114/mo
Claude API
Model: Anthropic Claude Sonnet 5.5 · Intelligence Index 54
Reliable at sticking to the given passages and citing them
anthropic.com ↗LLM observability
$29/mo
Langfuse · Core
Traces plus evals and datasets to measure answer quality
langfuse.com ↗All costs are estimates. Prices are read from each vendor's pricing page and checked every 12 hours (last: Oct 30, 2026); your bill depends on your traffic.
Model quality scores: Intelligence Index by Artificial Analysis.
The usage these costs assume
"Half" and "double" in the ranges scale these numbers.
- Monthly active usersA small team or early customer base
- 200 users
- Questions per userAbout two questions per working day
- 40 questions/user/month
- New documentsOngoing uploads after the initial backfill
- 300 documents/month
- Pages per documentTypical contract, report or policy length
- 15 pages
- File sizeText PDFs with some images
- 2 MB/document
- Documents keptStorage and index size grow with history
- 12 months
Engineering assumptions (17)
- Tokens per pageDense text page; scanned pages yield fewer after OCR
- 500 tokens
- Chunk sizeLarge enough to keep a clause or table together
- 800 tokens
- Chunks per answerEnough context for most questions without diluting it
- 6 chunks
- System prompt + questionInstructions, citation format and the user's question
- 600 tokens
- Answer lengthA few paragraphs with citations
- 350 tokens
- Tokens per questionA one-sentence question, embedded for search
- 30 tokens
- LlamaParse credits per pageBasic parsing is as low as 1 credit; layout-aware modes cost more
- 3 credits
- Stored size per chunk1,536-dim float vector (6 KB) plus text and metadata
- 8 KB
- Read units per searchA small index; bigger namespaces cost more per query
- 1 read units
- Stored size per question + answerQuestion, answer and citations
- 4 KB
- Cited PDFs openedSome answers send the user to the source page
- 0.3 opens/question
- Steps per ingestion jobParse, chunk, embed, index
- 4 steps
- Job time per pageMostly waiting on the parser and embedding APIs
- 2 seconds
- Trace units per questionTrace + retrieval + embedding + generation
- 4 units
- Team seatsPeople who look at traces
- 2 seats
- App vCPUA small Next.js server
- 0.5 vCPU
- App RAMA small Next.js server
- 0.5 GB
How to build it
- 1
Set up the app and database
Create the Next.js app, connect the database, add sign-in and a workspaces table so every document belongs to one workspace.
Uses: Web app & API, App database, AI framework
- 2
Upload to storage
Upload PDFs straight to object storage with signed URLs and record each file in the documents table.
Uses: File storage, App database
- 3
Ingestion pipeline
On upload, start a job: parse to text with page numbers, split into chunks with overlap, embed and upsert with workspace and page metadata. Show progress in the UI.
Uses: Ingestion jobs, PDF parsing, Embedding model, Vector database
- 4
Answering
Embed the question, fetch the top chunks for the user's workspace only, and stream an answer that cites document and page.
Uses: Embedding model, Vector database, Answer model, AI framework
- 5
Measure quality
Trace every question, collect thumbs up/down, and build a small evaluation set of real questions with expected sources.
Uses: LLM observability
- 6
Harden
Rate-limit questions per user, delete vectors when a document is deleted, and re-index when the parser or embedding model changes.
Uses: Web app & API, Vector database
Get this built for you
We'll adapt this blueprint to your product, set it up on your accounts and hand it over working, with the monthly bill you saw here.