AI engineering toolkit: split long documents into vector RAG chunks with overlap, calculate token costs against a dated model price registry, and scan prompts for common secret/PII patterns locally.
| Separator | Chunks | Avg len | Min len | Max len | Avg tokens (est.) | |
|---|---|---|---|---|---|---|
| {{ row.label }} | {{ row.count }} | {{ row.avgLen }} | {{ row.minLen }} | {{ row.maxLen }} | ~{{ row.avgTokensEst }} | Active |
Chunk counts and lengths change with the separator because each strategy splits the same text at different boundaries before the chunk size / overlap logic runs — a stat here is not a quality judgment on its own, just a shape comparison.
{{ c }}Sets Input Tokens below. Exact counts only exist for models with a public, vendorable tokenizer (currently OpenAI's o200k_base family) — everything else, including every non-OpenAI model here, is the chars÷4 heuristic because no public tokenizer exists to check it against.
This is a cost multiplier on token counts you provide — it does not tokenize text for you. Use the RAG Chunker's char÷4 estimate, or your provider's own tokenizer, for real prompts.
Paste either an OpenAI Usage API response ({"object":"page","data":[{"results":[...]}]}) or a flat array of {input_tokens, output_tokens, num_model_requests, model} records. Fills Input/Output Tokens with the per-call average and API Calls Count with the total — so the existing cost formula reproduces the real aggregate cost, not a copy of the raw totals.
{{ usageImportError }}
{{ usageImportSummary }}
| Model Name | Input Rate (/1M) | Output Rate (/1M) | Cost Per Call | Total Cost ({{ callsCount.toLocaleString() }} Calls) |
|---|---|---|---|---|
| {{ currentModel.display_name }} | ${{ currentModel.input_price }} | ${{ currentModel.output_price }} | ${{ perCallCost.toFixed(6) }} | ${{ totalCallsCost.toFixed(3) }} |
Source: {{ currentModel.source_url }} (verify before budgeting — prices change often)
{{ currentModel.notes }}
⚠ This price was only confirmed valid through {{ currentModel.price_valid_until }}, which has passed — re-check the source above before budgeting.
Reuses the model and per-call cost from above — {{ mau.toLocaleString() }} users × {{ callsPerUserMonth }} calls/mo = {{ (mau*callsPerUserMonth).toLocaleString() }} calls/mo.
| Line item | Monthly cost |
|---|---|
| LLM API ({{ currentModel ? currentModel.display_name : '—' }}) | ${{ monthlyLlmCost.toFixed(2) }} |
| {{ h.service.provider }} | ${{ h.cost.toFixed(2) }} |
| Total / month | ${{ totalMonthlyCost.toFixed(2) }} |
Reference estimate from list prices only — free tiers, discounts and taxes are not modeled. Vercel additionally meters Active CPU time and Provisioned Memory with no free tier, which this estimate does not include, so a real Vercel bill runs higher than shown here.
{{ sanitizedPiiText }}LLM Developer Workbench provides AI engineers and prompt developers with a workspace to split long documents into vector RAG chunks and compare chunking strategies side by side, count exact tokens for OpenAI models (chars÷4 estimate for everything else), import real OpenAI usage data to estimate cost, estimate token API costs against a dated multi-provider price registry, and scan for common secret/PII patterns locally.
Privacy: 100% Client-side. All document chunking, token estimation, and prompt sanitization run locally in your browser memory. Only the price registry JSON (public data, no user content) is fetched from this site's own server.
