Knowledge Base

The Knowledge Base enables Retrieval-Augmented Generation (RAG) — grounding AI responses in your organization's actual documents, policies, and data.

---

How It Works

flowchart LR
    subgraph Ingestion["Document Ingestion"]
        Upload["📄 Upload"] --> Chunk["✂️ Chunking"] --> Embed["🔢 Embedding"] --> Store["💾 Vector Store"]
    end

    subgraph Retrieval["Query-Time Retrieval"]
        Query["❓ User Question"] --> QEmbed["🔢 Embed Query"] --> Search["🔍 Similarity Search"] --> Context["📋 Top-K Context"]
    end

    Store -.-> Search
    Context --> LLM["🤖 LLM + Context"] --> Response["💬 Grounded Response"]

---

Document Ingestion

Supported Formats

Format	Description
PDF	Technical documents, policies, reports
Text	Plain text files, markdown
CSV	Structured data, catalogs, inventories

Ingestion Pipeline

Documents are processed through Apache Camel's integration engine:

1. Upload — via API or admin UI
2. Tenant isolation — documents are tagged with the tenant ID
3. Chunking — documents are split into manageable chunks (default: 1000 tokens with 200 token overlap)
4. Embedding — each chunk is converted to a vector using the configured embedding model
5. Storage — vectors are stored in MongoDB Atlas Vector Search

API

# Upload a document
POST /api/v1/kb/documents
Content-Type: multipart/form-data

file: @document.pdf
sourceType: MANUAL

---

Vector Search

MongoDB Atlas Vector Search

Synaptiq uses MongoDB Atlas Vector Search for semantic similarity:

• Index type — HNSW (Hierarchical Navigable Small World)
• Dimensions — 768 (nomic-embed-text) or 1536 (OpenAI)
• Similarity metric — Cosine similarity
• Top-K retrieval — configurable (default: 5 chunks)

Search API

# Semantic search
GET /api/v1/kb/search?q=refund+policy&limit=5

---

Embedding Models

Provider	Model	Dimensions	Notes
Ollama	nomic-embed-text	768	Local, free, recommended for dev
Google	Gemini Embedding	768	Cloud-based, production-ready
OpenAI	text-embedding-3-small	1536	Cloud-based, high quality

Configure in application-dev.yml:

spring:
  ai:
    ollama:
      embedding:
        model: nomic-embed-text

---

RAG in Chat

When a knowledge base is configured, chat responses are automatically augmented:

1. User's message is embedded
2. Top-K similar chunks are retrieved from the vector store
3. Retrieved context is injected into the system prompt
4. The LLM generates a response grounded in your documents
5. Source citations are included

RAG-Augmented Response

User: "What is our return policy for electronics?"

Response: Based on the Company Returns Policy (v2.3, Section 4.2), electronics can be returned within 30 days of purchase in original packaging with receipt. Opened items are subject to a 15% restocking fee. Defective items can be exchanged within 90 days under warranty.

Sources: Returns Policy v2.3 §4.2, Warranty Terms §2.1

---

Best Practices

Document Quality

• Use well-structured documents with clear headings and sections
• Avoid scanned PDFs without OCR — use text-based PDFs
• Keep documents up to date — outdated context leads to outdated answers

Chunk Size Optimization

• Smaller chunks (500 tokens) — better for precise Q&A
• Larger chunks (2000 tokens) — better for context-heavy responses
• Adjust overlap to prevent losing context at chunk boundaries

Domain-Specific Vocabularies

• Upload glossaries and terminology guides
• The RAG pipeline will ground responses in your domain's language