Max Mode (MASS-RAG)

Max Mode is Pegasus's advanced retrieval-and-reasoning pipeline. Instead of one LLM call against retrieved chunks, Max Mode runs a cross-encoder reranker plus three parallel filter agents and one synthesis agent. The result is more accurate answers on complex, multi-section, or contradictory documents — at the cost of more tokens and higher latency.

When to enable Max Mode#

Use Max Mode when:

• Questions require synthesis across multiple sections of a document.
• The knowledge base contains contradictory or conflicting information.
• The bot must perform multi-hop inference (A → B → C → conclusion).

Stick with Standard RAG when:

• Questions are simple and factual (answered in 1–2 paragraphs).
• Latency and cost are top priorities.
• Documents are short and low-noise.

How Max Mode works#

Vector retrieve (top 30)
         ↓
Cross-encoder reranker → top K
         ↓
3 filter agents in parallel:
  - Summarizer (high-level overview)
  - Extractor (verbatim quotes — anti-hallucination guardrail)
  - Reasoner (cross-section inference)
         ↓
Synthesis agent → final answer

Cross-encoder reranker#

Vector search is fast but ranks chunks by embedding similarity — it can promote a table of contents over the dense paragraph that actually answers the question. The reranker reads the query and each candidate chunk together with a cross-encoder model, producing a sharper relevance score. The top K chunks are forwarded; the rest are dropped before the LLM ever sees them.

Benefit: dramatically higher precision in the context window.

Trade-off: adds one extra service call per request. If the rerank service is unavailable, the system falls back to vector ordering — chat keeps working.

MASS-RAG multi-agent#

Three filter agents process the same context independently:

• Summarizer — produces a high-level overview of the retrieved context.
• Extractor — pulls verbatim quotes to anchor the answer in source material (anti-hallucination guardrail).
• Reasoner — performs cross-section inference and identifies relationships.

The synthesis agent never sees raw chunks — only the distilled outputs from the three filter agents. It combines them into the final answer.

Benefit: less "lost in the middle" — noise is filtered before synthesis. Lower hallucination risk via verbatim extraction. Multi-hop reasoning across document sections. Better handling of contradictions.

Trade-off: ~5–7× more tokens per request (4 LLM calls instead of 1). Latency is 2–3× Standard RAG. Not ideal for simple, single-paragraph questions.

Enabling Max Mode on a bot#

1. Open Source settings on the bot

   The "Max Mode (MASS-RAG)" section is below the system prompt.

2. Toggle Max Mode on

   Switch the toggle to enabled. Toast confirms: "Max Mode enabled."

3. Use it in chat

   In the Chat Workspace input bar, you'll see a Max Mode toggle. Enable it per-conversation to use Max Mode for that conversation. See Max Mode in chat.

Max Mode toggle on bot Source settings

Cost awareness#

Because Max Mode costs ~5–7× more tokens per chat message, it drains your daily credit limit much faster. Plan accordingly: use it for the questions that benefit, not as a default.