AI Assistant & Chatbot Development (Chat, Voice & Advanced RAG)

How documents are split decides what the retriever can find. The strategy is chosen per document type, and often several are combined.

Fixed-size chunking

Splits text every N tokens with an overlap. Fast and predictable; a good baseline for uniform text.

Recursive character chunking

Splits on paragraphs, then sentences, then words until chunks fit, keeping natural boundaries intact.

Sentence-based chunking

Groups whole sentences so no chunk ends mid-thought.

Semantic chunking

Uses embeddings to cut where the topic changes, so each chunk is about one idea.

Document-structure chunking

Follows headings, sections, Markdown or HTML structure, and keeps tables whole.

Sliding window

Overlapping windows so facts that span a boundary are never lost.

Parent-child (small-to-big)

Small chunks are matched precisely, then the larger parent section is sent to the LLM for context.

Agentic chunking

An LLM decides chunk boundaries and writes a title and summary for each chunk.

Late chunking

Embeds the whole document first, then pools token embeddings per chunk so each chunk keeps document context.

Finding the right evidence is most of the accuracy. Several retrievers run together and their results are fused and reranked.

Dense search

Embedding similarity in a vector database (Qdrant, FAISS, ChromaDB, Pinecone): finds meaning, even with different wording.

Keyword (sparse) search

BM25 matching of exact terms: product codes, names and numbers that embeddings miss.

Hybrid search

Dense and keyword search run together, so both meaning and exact matches are covered.

Reciprocal Rank Fusion (RRF)

Merges ranked lists from each retriever by rank position, giving a robust combined ranking without score calibration.

Cross-encoder reranker

A reranking model (for example bge-reranker or Cohere Rerank) rescores the top candidates against the question.

Maximal Marginal Relevance (MMR)

Picks results that are relevant but not repetitive, so the context covers more ground.

Metadata filtering

Restricts search by department, date, language or access rights.

Query rewriting & multi-query

Rephrases or expands the question into several searches, including follow-up questions in a chat.

HyDE

Generates a hypothetical answer and searches with it, which helps short or vague questions.

Patterns that make the system check, correct or replace its own retrieval.

Corrective RAG (CRAG)

Grades retrieved documents; if they are weak, it rewrites the query or falls back to another source before answering.

Self-RAG

The model decides when to retrieve and critiques whether its answer is supported by the evidence.

Adaptive RAG

Routes each question to no retrieval, single-step or multi-step retrieval depending on its complexity.

Agentic RAG

An agent plans searches across several tools and sources and combines the results.

GraphRAG

Builds a knowledge graph of entities and relations for questions that connect many documents.

Vectorless RAG

Reasoning-based retrieval that navigates a document's table of contents or page index like a human expert, with no embeddings.

Multimodal RAG

Retrieves from images, scanned pages and tables alongside text.

Auto skill selection

The assistant picks the right skill or pipeline (search, calculator, form, handoff) for each request.

Real-time voice needs low latency and natural conversation, not just text-to-speech on a chatbot.

Speech-to-text (STT)

Streaming transcription that handles accents and Bangla/English code-switching.

Text-to-speech (TTS)

Natural voices streamed sentence by sentence so replies start quickly.

LiveKit Agents

Real-time audio infrastructure for web voice and telephony integrations.

Voice activity detection & barge-in

Detects when the user starts speaking and stops the assistant mid-sentence, like a real conversation.

Latency budget

STT, retrieval, LLM and TTS tuned together to keep response time conversational.

Call handoff

Transfers to a human agent with a summary when the assistant should not handle it.

Accuracy is measured, monitored and protected.

RAGAS metrics

Faithfulness, answer relevancy, context precision and context recall on a test set.

Golden question set

Real user questions with expected answers, re-run after every change.

Citations

Answers point to their sources so users and admins can verify them.

Guardrails

Topic limits, refusal rules, PII masking and prompt-injection defenses.

Langfuse tracing

Every retrieval, prompt and response traced for debugging and quality review.

Grafana dashboards

Usage, latency and accuracy trends from PostgreSQL in real time.