Skip to main content
The knowledge base is what makes your chatbot smart. It contains all the content your chatbot can reference when answering questions. Without a knowledge base, your chatbot only has the AI model’s general knowledge.

​
How It Works

1

You add content

Upload documents, scrape websites, connect integrations, or type text manually.
2

FIFE.BOT processes it

Content is split into chunks, converted to vector embeddings (using OpenAI’s text-embedding-3-large), and stored in a PostgreSQL database with pgvector.
3

User asks a question

The question is analyzed, decomposed into sub-queries if needed, and searched against the knowledge base using hybrid search — combining vector similarity and full-text search.
4

Best chunks are retrieved

Results are fused, deduplicated, diversified by source, reranked by the LLM, and checked for coverage. If coverage is insufficient, additional retrieval rounds run automatically.
5

AI generates a response

The selected chunks are passed as context to the AI model, which generates a grounded response with source citations.

​
Source Types

​
The Knowledge Base Tab

Open any chatbot and click the Knowledge Base tab to see:
  • Add Source dropdown — pick a source type to add
  • Filter bar — filter by source type (All, Websites, Documents, Texts, Table, Notion, Confluence, Google Drive, SharePoint)
  • Source list — all added sources with status indicators
  • Processing stats — total sources, ready count, processing count

​
Source Status Indicators

​
Per-Source Actions

​
Routing Instructions

Each knowledge base source can have optional routing instructions — extra context that tells the AI how to use content from that specific source. Example: A website source for your pricing page might have routing instructions like:
“When answering pricing questions from this source, always mention the 14-day free trial and the annual discount.”
Routing instructions are combined with the system prompt at query time.

​
Plan Limits

Limits match fife.bot pricing — your plan caps chatbots, monthly credits, and knowledge sources (each website, document set, integration, etc. counts as a source).

​
Search Technology

FIFE.BOT uses a sophisticated retrieval pipeline:
  1. Query analysis — determines intent, language, and complexity
  2. Multi-query decomposition — complex questions are split into sub-queries
  3. Hybrid search — vector cosine similarity + PostgreSQL full-text search with weighted scoring
  4. Reciprocal Rank Fusion — merges results from multiple search strategies
  5. Source diversification — ensures results come from multiple sources
  6. Parent chunk expansion — retrieves surrounding context for better answers
  7. LLM reranking — the AI re-scores chunks for relevance
  8. Coverage checking — if the answer isn’t fully covered, additional retrieval rounds run
This means your chatbot doesn’t just do keyword matching — it understands the meaning of questions and finds the most relevant content, even if the exact words don’t appear in your documents.