Analyzer Agent

Role: Relevance Scoring and Reranking Specialist Location: Inline in api/routes/search.js (analyzer prompt)

Overview

The Analyzer processes federated search results from the Researcher agent. It scores each result by relevance using the research plan as context, filters low-scoring results (below 60), extracts key sentences, highlights core terms, and generates brief answers.

Scoring Criteria

Results are scored 0-100 based on four weighted factors:

Criterion	Weight	Description
Direct Relevance	40%	How directly does the passage address the user’s query?
Depth of Insight	30%	Substantive teaching vs surface mention?
Research Plan Alignment	20%	Does it address angles/facets from the research plan?
Unexpectedness	10%	Does it challenge assumptions or reveal surprising perspectives?

Analysis Tasks

1. Score passages (0-100) using the weighted criteria above
2. Filter passages scoring below 60
3. Extract the most relevant sentence (using start/end anchors)
4. Identify 3-7 core terms to highlight within that sentence
5. Summarize with a brief 5-8 word answer from the quote’s perspective
6. Explain why the passage is relevant (context of research plan)

Output Format

The analyzer returns JSON with ranked results (highest score first):

{
  results: [
    {
      originalIndex: 0,           // Index in original search results
      id: "passage_123",          // Passage ID for tracking
      score: 95,                  // Relevance score (0-100)
      briefAnswer: "Divine love unites all faiths",  // 5-8 word answer
      sentenceStart: "In the sight of",  // First 3-5 words VERBATIM
      sentenceEnd: "one human family.",  // Last 3-5 words VERBATIM
      coreTerms: ["divine", "love", "unites", "human", "family"],
      relevanceNote: "Addresses cross-traditional unity angle from research plan"
    }
  ],
  introduction: "Brief 1-2 sentence intro orienting the user to what was found.",
  semanticNote: "Optional: note if key terms appear in distinctly different senses across passages"
}

Brief Answer Rules

The brief answer must:

• Answer the query directly from the quote’s perspective
• Be 5-8 words maximum
• NO meta-language (“this passage states”, “asserts”, “discusses”)
• Just the answer itself

Good: “Divine love unites all faiths” Bad: “This passage asserts that divine love unites faiths”

Sentence Anchoring

Instead of copying entire sentences (which can fail due to LLM rephrasing), the analyzer uses anchor-based matching:

1. sentenceStart: First 3-5 words VERBATIM from the text
2. sentenceEnd: Last 3-5 words VERBATIM (including punctuation)

The route handler uses these anchors to find and highlight the exact sentence in the original text.

Semantic Awareness

The analyzer notes when key terms appear with different philosophical meanings across passages:

• “equality” → rights vs outcomes vs nature vs dignity
• “freedom” → from constraint vs to act vs spiritual vs political
• “justice” → retributive vs restorative vs distributive vs divine
• “love” → divine vs human vs duty vs emotion

These distinctions are captured in semanticNote when relevant.

Research Plan Context

The analyzer receives the full research plan from the Researcher agent:

RESEARCH STRATEGY:
- Reasoning: [Overall search strategy explanation]
- Assumptions being challenged: [List of expected answers to disprove]
- Search queries used:
  1. "query text" (hybrid) - [rationale]
  2. "query text" (semantic) - [rationale]
- Traditions covered: [List of religions/traditions]
- Surprises to watch for: [Unexpected findings to highlight]

This context helps the analyzer:

• Understand why each passage was found
• Score passages that address multiple research angles higher
• Identify passages that challenge stated assumptions
• Note which tradition/angle each result addresses

Sorting

The analyzer ranks results by score in its JSON output. The route handler sorts the enhanced sources by score before sending to the frontend, ensuring results are always ordered by relevance.

Highlighting Flow

1. Analyzer identifies relevant sentence via anchors
2. Route handler matches anchors in original text
3. Core terms within the sentence get <b> tags
4. Entire sentence gets <mark> tags
5. Result: ${before}<mark>The <b>divine</b> <b>love</b>...</mark>${after}

Configuration

The analyzer runs inline in the search route with:

await ai.chat([
  { role: 'system', content: 'You are an expert search result analyzer. Return only valid JSON, no markdown.' },
  { role: 'user', content: analyzerPrompt }
], {
  temperature: 0.3,    // Low for consistent scoring
  maxTokens: 3000      // Higher for processing multiple results
});

Error Handling

If analysis fails, the route falls back to sending original results with:

• Empty summaries
• Plain text (no highlighting)
• No scoring/ranking