The words we use, in plain English.

A

AEO and GEO

Answer engine optimisation (AEO) and generative engine optimisation (GEO) are practices that make content easy for AI assistants and AI search, such as ChatGPT, Perplexity and Google AI Overviews, to understand, quote and cite.

Web application development

Agentic AI

Agentic AI describes AI systems that plan and carry out multi-step tasks with some independence, using a language model to decide which tool or action comes next, rather than producing a single response to a single prompt.

We set the level of autonomy step by step. Reading and drafting can run on their own; anything expensive or irreversible waits for a person.

Custom AI agents, RAG and private LLMs

AI agent

An AI agent is software that uses a language model to plan and carry out multi-step tasks by calling tools, such as searching a database, updating a record or sending a draft for approval, rather than only answering a question.

We give agents a fixed set of tools, test them against real scenarios, and put human approval on any step where a mistake is expensive.

Custom AI agents, RAG and private LLMs

AI copilot

An AI copilot is an assistant embedded in a tool people already use, which drafts or suggests actions that a person reviews and approves before anything happens.

We build a copilot when a person should approve each step. When the work spans several steps and systems, it may be an agent instead.

AI product development

AI-native

AI-native describes software, or a company, designed with language models as a core part from the start, rather than added later as a feature.

For us it means AI products measured with an evaluation set before launch, and engineering that uses AI assistants with every change reviewed by a person.

About ByteKnot Labs

AI-native CRM

An AI-native CRM is a customer relationship management system with a language model built into how it captures and uses information: it summarises calls and emails, drafts follow-ups, scores leads from your own history and flags exceptions.

AI-native CRM vs traditional CRM

C

Chunking

Chunking is splitting documents into smaller passages before indexing them for retrieval. Chunk size and boundaries decide what the model can be shown, so they have a large effect on RAG quality.

Code-switching

Code-switching is moving between two languages within one conversation or one sentence, as in Hinglish, which mixes Hindi and English. Speech recognition then has to handle both languages, and both scripts, in the same line.

AksharSTT: Hinglish speech to text

Context window

A context window is the maximum amount of text, measured in tokens, that a language model can take in and produce in one request, including instructions, retrieved passages and the answer.

RAG vs fine-tuning

E

Embeddings

An embedding is a list of numbers that represents the meaning of a piece of text, an image or other data, so that items with similar meanings sit close together and can be found by similarity search.

Evaluation set

An evaluation set, also called an eval set, test set or golden set, is a collection of real inputs with the outputs a team considers correct, used to score an AI system automatically every time its prompt, model or data changes.

We build it from your real cases before writing the first prompt. It turns 'it seems better' into a number.

F

Fine-tuning

Fine-tuning is further training of an existing language model on your own examples, so it adopts a particular style, format or behaviour more consistently than prompting alone achieves.

We fine-tune with LoRA or QLoRA in your account, and only when the evaluation set shows that prompting and retrieval are not enough.

RAG vs fine-tuning

G

Guardrails

Guardrails are checks placed around a language model to block or correct unwanted inputs and outputs, such as off-topic requests, leaked personal data or answers in the wrong format.

Constrained outputs, checks on inputs and outputs, and approval gates on expensive actions.

H

Hallucination

A hallucination is a fluent, confident answer from a language model that is false or unsupported by its sources.

Retrieval with citations makes hallucinations visible; evaluation sets measure how often they happen.

Human-in-the-loop

Human-in-the-loop is a design in which a person reviews, approves or corrects an AI system's output at set points before it takes effect, such as before an email is sent or a record is changed.

We call these approval gates and put them wherever a wrong move would be expensive.

Hybrid search

Hybrid search combines keyword search, which matches exact terms such as product codes and names, with semantic search over embeddings, which matches meaning, and merges the results.

I

Infrastructure as code

Infrastructure as code means describing servers, networks, databases and permissions in version-controlled files, using tools such as Terraform, so environments can be reviewed, reproduced and rebuilt automatically.

K

Knowledge distillation

Knowledge distillation trains one model, the student, to reproduce the outputs of another, the teacher, so the student learns from the teacher's predictions as well as from labelled data. It is a common way to make a model smaller or to teach it a new language.

How AksharSTT was trained

L

Large language model (LLM)

A large language model is a neural network trained on large amounts of text to predict and generate language, used for answering questions, summarising, drafting, classifying and calling tools.

LLM-as-a-judge

LLM-as-a-judge is an evaluation method in which one language model scores another's output against criteria such as correctness, faithfulness to sources or tone, so large evaluation sets can be scored automatically. Its scores are only useful once they have been checked against human ratings on a sample.

LLMOps

LLMOps is the practice of running language-model applications in production: versioning prompts and models, tracing every call, watching cost, latency and quality, and re-running evaluations whenever something changes.

LoRA and QLoRA

LoRA (low-rank adaptation) is a fine-tuning method that trains a small set of extra weights instead of the whole model, which makes fine-tuning far cheaper. QLoRA does the same on a compressed model to save more memory.

M

Model card

A model card is a short document that records what a model is for, what data it was trained or tuned on, how it was evaluated, its known limits and who owns it.

Every custom model we hand over comes with one, for audits.

Model Context Protocol (MCP)

The Model Context Protocol is an open standard for connecting AI models to tools and data sources through a common interface, so one integration works across different models and clients. Since December 2025 it has been governed by the Agentic AI Foundation, a directed fund of the Linux Foundation.

We put an MCP server in front of your systems so agents reach them through one audited door.

O

Open-weight model

An open-weight model is a language model whose trained weights are published, so it can be downloaded, fine-tuned and run on your own hardware. Llama, Mistral, Qwen and Gemma are examples.

P

Private LLM

A private LLM is a language model deployed on infrastructure an organisation controls, such as its own GPUs or a private cloud account, so prompts and data do not pass through a third-party model API.

We serve open-weight models with vLLM on your hardware, OCI or Google Cloud when data cannot leave your network.

Fine-tuning and private models

Prompt injection

Prompt injection is an attack in which instructions hidden in user input, a document, an email or a web page trick a language model into ignoring its original instructions, for example to reveal data or misuse a tool.

We treat retrieved content as untrusted, limit what each tool can do, and include injection attempts in red-team checks.

Q

Quantisation

Quantisation stores a model's weights in fewer bits, such as 4-bit instead of 16- or 32-bit numbers, so the model needs less memory and runs faster on ordinary hardware, usually at a small cost in accuracy. Quantisation-aware training keeps that cost down by training at the lower precision from the start.

AksharSTT was trained with 4-bit weights from its first update, which is how a 55-million-parameter model fits in 29.5 MB.

AksharSTT size and speed

R

Red-teaming

Red-teaming is deliberately trying to make an AI system misbehave, through tricky, adversarial or malicious inputs, to find weaknesses before real users do.

Reranking

Reranking is a second retrieval step in which a model re-orders the top search results by how well each one answers the question, so the passages that answer it reach the language model first.

Retrieval-augmented generation (RAG)

Retrieval-augmented generation is a technique where a system first searches your documents or data for relevant passages, then gives them to a language model to write an answer, usually with citations to the sources.

RAG systems

S

Sideloading

Sideloading is installing an app on a device from outside the official app store. On iPhone, apps signed with a free Apple ID expire after 7 days unless they are re-signed.

ByteLoader

Speech to text (STT)

Speech to text, also called automatic speech recognition (ASR), is software that turns spoken audio into written text. A model listens to the recording and writes down the words, often with a time stamp and a confidence score for each one.

Our AI Lab publishes AksharSTT, speech to text for Hindi, English and Hinglish that runs offline on an ordinary CPU.

AksharSTT

Spring Boot

Spring Boot is a Java framework for building production-ready web services and applications with sensible defaults for security, data access, configuration and monitoring.

Spring Boot vs Node.js

System of record

A system of record is the authoritative source for a type of business data, such as customers in a CRM or orders in an ERP, which other systems read from and trust.

T

Tool calling

Tool calling, also called function calling, is a language model's ability to return a structured request to run a specific function, such as a database query or an API call, with the arguments filled in, which the application then runs.

V

Vector database

A vector database stores numerical representations of text, images or other data (embeddings) and finds the items most similar in meaning to a query. It is the search layer behind most RAG systems.

We use Qdrant when filtering, hybrid search or scale demand it, and the VECTOR type in MySQL 9 or pgvector when the relational store is enough.

W

Word error rate (WER)

Word error rate is the standard accuracy measure for speech to text: the words a model substitutes, drops or adds, divided by the number of words in the correct transcript. Lower is better, and two models are only comparable when they are scored on the same test set with the same text normalisation.

We publish word error rate for each public test set, beside a named reference model, including the sets where our model loses.

AksharSTT accuracy

Z

Zero data retention

Zero data retention is an arrangement under which a model provider does not store prompts and outputs after returning a response. Each provider defines and offers it differently, so the terms are checked provider by provider.