Chapter 6 — KnowledgeBase

Overview & Architecture

Every VoiceBot is backed by one or more Topics. Each Topic maps to an isolated RAG server instance holding curated documents, an auto-generated knowledge graph and a tunable retrieval pipeline.

Figure 1 — KnowledgeBase management administration console.

What administrators do here

1. Organise knowledge into thematic Topics — for example “Product Support”, “Sales & Marketing”, “HR & Onboarding”.
2. Upload documents that are automatically chunked, embedded and indexed.
3. Query the knowledge base using five supported search modes.
4. Visualise the knowledge graph to understand entity relationships extracted from documents.
5. Configure and monitor the underlying RAG server instances.

Supported document types

TXT MD DOCX PDF PPTX XLSX RTF ODT EPUB HTML / HTM TEX JSON XML YAML / YML CSV LOG CONF SQL BAT SH C H CPP PY JAVA

The list spans office and text formats, e-books and markup, structured data and configuration files, and source code in C, C++, Python and Java.

Key features

Feature	Description
Multi-topic management	Create, edit and delete Topics, each with its own colour, tags and dedicated RAG server.
Drag & drop upload	Upload documents by drag-and-drop or file picker, with real-time progress and chunking and indexing feedback.
Document preview	View document metadata — size, chunks, status, upload date — and content directly in the browser.
Multi-mode query interface	Query using five search modes (Naive, Local, Global, Hybrid, Mix) with configurable Top-K, temperature and context-only options.
Interactive knowledge graph	Canvas-based visualiser with zoom, pan, drag, search and type filtering, showing entities, relations, chunks and their connections.
Server health monitoring	Real-time health checks with online, offline and connecting indicators on every Topic card and in the sidebar.
Hybrid API client	Attempts real HTTP calls to RAG servers and falls back automatically on failure, so the interface keeps working.
Global configuration	Centralised settings for API keys, default ports, health check intervals, security and logging.
RAG tuning	Per-topic configuration of LLM model, embedding model, chunk size, chunk overlap, max tokens, temperature, Top-K, graph community level and graph search toggle.

Topic-per-server model

Every Topic maps to an independent RAG server instance. This one-to-one mapping provides:

• Data isolation — different Topics can use different LLM and embedding models optimised for specific languages.
• Domain separation — documents, graph entities and vector indices are fully isolated across Topics, with no cross-contamination.
• Independent scaling — each server is sized, deployed and restarted according to its own document volume and query load.
• Operational independence — a failure or reindex on one Topic does not affect any other.

Figure 2 — KnowledgeBase data and language isolation.

Text ingestion pipeline

Uploaded documents pass through eight stages before becoming queryable.

Step 1

Document upload & validation

The file is uploaded through the drag-and-drop interface or file picker. The server validates type and size and acknowledges receipt while the interface shows real-time progress.

Step 2

Text extraction & normalisation

The raw file is parsed into plain text — OCR or text-layer extraction for PDFs, XML unpacking for DOCX, direct reads for Markdown and plain text — then normalised for whitespace, encoding and special characters.

Step 3

Chunking with overlap

The text is split into chunks governed by two parameters: chunk size (default 1200 characters) and chunk overlap (default 100 characters), so ideas spanning a boundary are not lost.

Step 4

Vector embedding generation

Each chunk passes through an embedding model to produce a high-dimensional vector capturing its semantic meaning, enabling similarity search.

Step 5

Vector store indexing

Vectors, with references back to their source chunks and parent documents, are inserted into a vector store optimised for approximate nearest-neighbour lookups.

Step 6

Entity & relationship extraction

An LLM scans each chunk to identify named entities — people, products, concepts, technical terms — and the relationships between them, returning structured (subject, predicate, object) triples.

Step 7

Knowledge graph construction

Entities become nodes and relationships become edges in a persistent graph. Duplicates are merged and a community detection algorithm groups related nodes into hierarchical communities (default depth 2); each community receives a generated summary that powers the Global and Mix modes.

Step 8

Status update & availability

Once both the vector index and the graph are populated, the document's status moves from processing to indexed and it becomes fully queryable.

Figure 3 — Text ingestion into the KnowledgeBase pipeline.

Configuration influence

Every RAG parameter affects this pipeline directly. Changing chunkSize alters how granular the splits are; changing embeddingModel changes the vector space; changing graphCommunityLevel controls how many layers of community summaries are built. All parameters can be adjusted per Topic from the Settings tab before uploading documents.

Retrieval and answer generation

A query follows one of five retrieval strategies, selected in the Search Mode dropdown. Each determines what is retrieved before the LLM generates its answer.

Mode	Retrieved	How it works
Naive	Top-K most similar chunks from the vector store	The query is embedded and compared against all chunk vectors by cosine similarity; the closest chunks become the context. Fast and straightforward, best for factual lookups.
Local	Graph neighbourhood around matched entities	Entities mentioned in the query are located in the knowledge graph, and their immediate neighbourhood is traversed to gather structured context about those entities.
Global	Community-level summaries from the graph	Pre-computed summaries of entire communities are retrieved, providing broad thematic context that captures high-level patterns across many documents.
Hybrid	Chunks plus local graph neighbourhood	Combines Naive and Local: chunk similarity provides textual evidence while graph traversal adds entity context. The merged context is deduplicated and ranked. Recommended default.
Mix	Local neighbourhood plus global summaries	Combines Local and Global, capturing both fine-grained entity detail and broad thematic patterns — ideal for complex questions spanning multiple concepts.

After retrieval, the context fragments are assembled into a single prompt and sent to the configured LLM together with the original query. The answer is returned to the interface with metadata: source document names, token usage, processing time and the search mode used.