Knowledge Base

Knowledge sources give the AI relevant context for each query, grounding responses in your actual content.

FAQ Knowledge Source

Best for structured question-answer pairs:

import { FAQKnowledgeSource } from '@chatcops/core/knowledge';

const faq = new FAQKnowledgeSource([
  { question: 'What are your business hours?', answer: '9am-5pm EST, Monday to Friday.' },
  { question: 'Do you offer refunds?', answer: 'Yes, full refunds within 30 days.' },
  { question: 'How do I contact support?', answer: 'Email support@example.com or use this chat!' },
  { question: 'What payment methods do you accept?', answer: 'Visa, Mastercard, PayPal, and wire transfer.' },
]);

How FAQ Matching Works

1. User’s query is tokenized into words
2. Each FAQ pair is scored by word frequency overlap (question + answer)
3. Top 5 matching FAQs are returned as context
4. Context is formatted as Q: ...\nA: ... and injected into the AI prompt

Text Knowledge Source

Best for unstructured content (documentation, product descriptions, policies):

import { TextKnowledgeSource } from '@chatcops/core/knowledge';

const docs = new TextKnowledgeSource(
  `Our company was founded in 2020 by Jane Doe.
  We specialize in AI-powered customer support solutions.
  Our flagship product is ChatCops, an embeddable chatbot widget.
  We serve over 500 businesses worldwide.
  Our headquarters is in San Francisco, CA.`,
  { chunkSize: 500 }  // Optional: characters per chunk
);

How Text Matching Works

1. Source text is split into chunks at initialization
2. User’s query words are matched against each chunk
3. Top 3 relevant chunks are returned
4. Context is prefixed with "Relevant context:\n"

KnowledgeSource Interface

interface KnowledgeSource {
  type: string;
  getContext(query: string): Promise<string>;
}

Multiple Sources

Combine multiple knowledge sources — all contexts are concatenated:

chatcopsMiddleware({
  provider: { type: 'claude', apiKey: '...' },
  systemPrompt: 'Answer using the provided context. If unsure, say so.',
  knowledge: [
    new FAQKnowledgeSource(faqs),
    new TextKnowledgeSource(productDocs),
    new TextKnowledgeSource(policyDocs),
  ],
});

Custom Knowledge Source

Implement KnowledgeSource for database-backed or vector search:

class VectorKnowledgeSource implements KnowledgeSource {
  type = 'vector';

  async getContext(query: string): Promise<string> {
    const results = await vectorDB.search(query, { topK: 5 });
    return results.map(r => r.text).join('\n\n');
  }
}