How do you build Bengali AI apps with Plugsky?

Key facts

Bengali script	Left-to-right abugida; conjuncts and matras need NFC normalisation
Tokenisation	Conjuncts and vowel signs fragment into subwords — measure with the Plugsky token calculator
Code-mixing	Bengali script and Romanised Banglish can share one collection via multilingual embeddings
API compatibility	OpenAI-compatible POST https://api.plugsky.com/v1/chat/completions — chat, streaming, JSON mode and function calling
Models	30+ models behind one endpoint, from free tiers to frontier reasoning
Free tier	Free plan with plugsky-micro and plugsky-lite, no card required
Deployment	Plugsky cloud, your VPC, on-prem and air-gapped options
Product status	Chat, streaming, JSON mode, function calling, embeddings, RAG and agents live; audio, images, batch and fine-tuning coming soon

TL;DR

• One base_url change serves Bengali script and Banglish text.
• NFC-normalise matras and conjuncts before indexing or prompting.
• Measure tokens on real text — character counts underestimate.
• plugsky-embed-multilingual supports mixed Bengali/English retrieval.
• Evaluate with a native speaker; conjunct errors are easy to miss.

How it works, step by step

1. Create a Plugsky API key on the free plan (no card) and set base_url to https://api.plugsky.com/v1.
2. Send a small set of real Bengali prompts to /v1/chat/completions and compare output across two or three models.
3. Count tokens for those prompts with the Plugsky token calculator and set chunk sizes that fit your model context.
4. Normalise text before indexing or prompting: NFC-normalise the text and keep numerals consistent between indexing and queries.
5. For RAG, embed with plugsky-embed-multilingual and test cross-language queries alongside Bengali-only queries.
6. Score candidate models on a Bengali gold set with native-speaker review, then cut production traffic over.

Try it yourself

Open the token calculator →

How Plugsky handles Bengali text

Bengali is written left-to-right in the Bengali abugida, where consonant clusters (conjuncts) and vowel signs (matras) attach to a base character.

Kolkata and Dhaka orthography differ in spelling and vocabulary, and everyday writing mixes English words freely — the register usually called Banglish.

Have a Bengali speaker review register and spelling on a small gold set; automated metrics rarely catch conjunct and matra errors that readers notice immediately.

Tokenisation and cost in Bengali

Bengali is character-dense and tokenisers fragment conjuncts and matras, so token counts can exceed what the word count suggests. Always measure real text instead of estimating from character counts.

• Normalise Unicode to NFC so matras and conjuncts stay composed.
• Test Bengali script and Romanised Banglish separately — they tokenise very differently.
• Keep numerals consistent (Bengali or Western) in prompts and retrieval.
• Chunk on sentence boundaries; Bengali sentences are long and punctuation-light.

Bengali retrieval and RAG

Index Bengali documents as they are written, but also test Roman-script queries, because many users search in Banglish even when the corpus is in Bengali script.

• Use plugsky-embed-multilingual for Bengali and mixed Bengali/English corpora.
• NFC-normalise text and strip zero-width joiners that do not change meaning.
• Store a transliterated variant if users query in Roman script.
• Evaluate retrieval with both script and Roman queries.

Code example: a Bengali request

Point your existing OpenAI client at https://api.plugsky.com/v1 and pass Bengali text in the content field — no language flag and no separate endpoint. Streaming, JSON mode and function calling keep the same request shapes.

client = OpenAI(base_url="https://api.plugsky.com/v1", api_key=os.environ["PLUGSKY_API_KEY"])

client.chat.completions.create(model="plugsky-pro", messages=[{"role": "user", "content": "এই চুক্তির তিনটি মূল পয়েন্ট বাংলায় সংক্ষেপে লেখো।"}])

Start on the free plan with plugsky-micro and plugsky-lite, then compare paid models on a Bengali gold set before cutover. See the docs for request details.

For production, log the model name and your normalisation settings with each request, and re-run the Bengali gold set whenever either changes — language quality regressions usually come from prompt or preprocessing drift, not from the model alone.

Honest comparison

Capability	Plugsky	Bengali workflow today	Building in-house
API compatibility	OpenAI-compatible — change base_url and model name	Varies by provider and SDK	Full rewrite
Bengali text handling	NFC-normalised script plus Banglish query support	Depends on provider tokeniser and prompt hygiene	You build normalisation, segmentation and evals
Token budget	Fixed tokeniser per model; measure with the Plugsky token calculator and chunk to fit	Varies by provider and model	You host and tune each tokeniser
Multilingual retrieval	plugsky-embed-multilingual available for cross-language RAG	Often needs a separate embedding vendor	You serve and maintain embeddings
Sovereignty	Cloud, VPC, on-prem and air-gapped with residency options	Usually US/EU public endpoints	You own the full stack

Frequently asked questions

Can Plugsky handle Bengali text?

Yes. The API accepts UTF-8 Bengali input on the OpenAI-compatible chat endpoint; output quality depends on the model, so compare two or three on your own prompts before choosing.

How do I estimate token usage for Bengali?

Bengali is character-dense and tokenisers fragment conjuncts and matras, so token counts can exceed what the word count suggests. Use the token calculator at /tools/llm-token-calculator before sizing context windows or chunk lengths.

Should I send Bengali script or Banglish?

Both work. Script is better for formal content; many users type Banglish, so support Roman queries with transliteration or test them against plugsky-embed-multilingual.

Do I need to handle conjuncts manually?

No — normalise Unicode to NFC and conjuncts and matras stay composed. Test a sample of real text to catch mixed normalisation forms.

Is there a multilingual embedding model?

Yes — plugsky-embed-multilingual is part of the 30+ model catalogue and is built for cross-language retrieval. Keep one embedding model per vector collection.

Can I keep data in my region?

Plugsky supports cloud, VPC, on-prem and air-gapped deployment with data-residency options; confirm your requirements with the docs and the enterprise team.

How do I migrate an existing app?

Change base_url to https://api.plugsky.com/v1 and map the model name. Streaming, JSON mode, function calling and embeddings keep the same request shapes.

Is there a free plan?

Yes — the free plan includes two free models, plugsky-micro and plugsky-lite, with no card. A 14-day full-access trial unlocks the paid catalogue.