Best LLMs for Content Set Relevance Scoring
Scores every item in one supplied set of retrieved content for relevance to a query or analysis objective, returning a calibrated item-level score with concise evidence for each. Illustrative uses include scoring support articles against a ticket, procurement documents against an
Models
Frontier on this task: Gemini 3.5 Flash at 7.05 / 10. Quality bar at 90%: 6.35.
point-estimate floor (CI low) · upper CI (less certain) · Bars sorted by blended cost; best-value model first. Greyed rows are MEDIUM+ models whose point estimate clears the bar but whose CI low does not.
| Model | Quality score | CI low | Cost / 1k runs | vs best value |
|---|---|---|---|---|
| GPT-5.6 Luna | 6.36 / 10 | 6.01 | $6.86 | best value |
| Thinking Machines Inkling Small | 6.68 / 10 | 6.32 | $15.76 | 2.3x more expensive |
| GPT-5.6 Terra | 7.01 / 10 | 6.72 | $39.89 | 5.8x more expensive |
| Gemini 3.5 Flash | 7.05 / 10 | 6.76 | $41.26 | 6x more expensive |
| GPT-5.6 Sol | 6.91 / 10 | 6.54 | $65.50 | 9.5x more expensive |
| Claude Haiku 4.5 | 5.03 / 10 | 4.76 | $25.55 | 3.7x more expensive |
| DeepSeek V4 Pro | 5.63 / 10 | 5.39 | $71.00 | 10x more expensive |
| GPT-5.4 Nano | 3.43 / 10 | 3.05 | $4.92 | 28% cheaper |
| MiniMax M3 | 5.47 / 10 | 5.19 | $15.24 | 2.2x more expensive |
| Gemini 3.5 Flash Lite | 5.68 / 10 | 5.19 | $7.57 | 1.1x more expensive |
| NVIDIA Nemotron-3 Nano 30B-A3B | 2.50 / 10 | 2.14 | $4.21 | 39% cheaper |
| NVIDIA Nemotron-3 Super 120B | 5.23 / 10 | 4.74 | $19.87 | 2.9x more expensive |
| Qwen 3.7 Plus | 5.63 / 10 | 5.17 | $22.79 | 3.3x more expensive |
| Tencent Hy3 | 5.85 / 10 | 5.40 | $10.08 | 1.5x more expensive |
| Gemini 3.1 Flash Lite | 4.95 / 10 | 4.66 | $5.80 | 15% cheaper |
| DeepSeek V4 Flash | 5.42 / 10 | 5.13 | $23.87 | 3.5x more expensive |
Cost breakdown
| Model | Quality | Confidence | Cost / 1k runs | Overpay | Mode |
|---|---|---|---|---|---|
| GPT-5.6 Luna ★ OpenAI | 6.36 / 10 CI [6.01, 6.72] | MEDIUM | $6.86 | best value | batch |
| Thinking Machines Inkling Small OpenRouter | 6.68 / 10 CI [6.32, 7.04] | MEDIUM | $15.76 | 2.3x | batch |
| GPT-5.6 Terra OpenAI | 7.01 / 10 CI [6.72, 7.30] | HIGH | $39.89 | 5.8x | batch |
| Gemini 3.5 Flash best Gemini | 7.05 / 10 CI [6.76, 7.35] | HIGH | $41.26 | 6x | batch |
| GPT-5.6 Sol OpenAI | 6.91 / 10 CI [6.54, 7.28] | MEDIUM | $65.50 | 9.5x | batch |
Overpay shows how much more you pay than the best-value model that clears the quality bar (marked ★) — the best-value good-enough option. "16x" means you overpay 16× — 16× that reference for no quality benefit above the bar. Typical call shape for this task: 15899 input tokens → 3765 output tokens, EMA-tracked from production traffic. Cost is the observed, all-in $ per 1,000 task runs: each model's own measured usage on this task — output verbosity, thinking/reasoning tokens, cache reads and writes, and the spend on its billed failures — priced at current list rates and adjusted by the billing overhead we actually reconcile against provider invoices. Models that answer tersely cost what they actually cost; models that think at length pay for it. Not comparable to providers' advertised $/1M list rates — this is what running the task costs, not a per-token price.
Evaluation rubric
Judge complete one-to-one coverage of the supplied set, item-level semantic relevance, score calibration and consistency across items, threshold correctness, evidence quality, treatment of partial relevance, preservation of item identity or order, and resistance to keyword/source-prestige shortcuts.
Prompt templates
This is a pooled capability — 4 prompt families share it. The pair shown first is the most frequently used in production.
LLMB_CONTENT_SET_RELEVANCE_SCORING_SYSTEM +
LLMB_CONTENT_SET_RELEVANCE_SCORING_USER
(2086 calls in window)
System prompt
Evaluate every item in the supplied content set for substantive relevance, not mere keyword overlap. Consider whether each item contains information that can advance the supplied query or objective, and distinguish direct evidence, useful context, passing mention, and irrelevance. Score each item independently based only on its supplied content; do not infer quality from source prestige alone. Apply the caller’s threshold consistently and return one result for every input item. Treat empty optional values as absent and return only the requested result. Your response must conform exactly to this output schema: {schema_json_string}.
User prompt
Inputs — content_type: {content_type}; query: {query}; posts: {posts}; relevance_threshold: {relevance_threshold}. Use only these inputs to complete the task defined by the system prompt.
POST_RELEVANCE_SYSTEM_PROMPT +
POST_RELEVANCE_USER_PROMPT
(144 calls in window)
System prompt
You are an expert content analyst specializing in relevance assessment for social media posts and other short-form content. Your task is to analyze a batch of posts and determine each item's relevance to a specific search query. The exact content type and the classification threshold are provided in the user message.
## Your Role:
- Evaluate each post's relevance to the given query with precision and consistency
- Assign accurate relevance scores on a 0.0 to 1.0 scale
- Focus on semantic relevance, not just keyword matching
## Relevance Scoring Guidelines:
**1.0 (Perfect Match)**:
- Directly addresses the exact query topic with substantial detail
- Contains primary keywords and concepts from the query
- Provides valuable information directly related to the query
**0.8-0.9 (Highly Relevant)**:
- Strongly related to the query topic
- Contains most key concepts from the query
- Provides useful information on the topic
**0.6-0.7 (Moderately Relevant)**:
- Related to the query but may be tangential
- Contains some relevant keywords or concepts
- Provides some useful context or related information
**0.4-0.5 (Somewhat Relevant)**:
- Loosely related to the query topic
- May contain peripheral concepts or broader category relevance
- Limited direct value for the specific query
**0.1-0.3 (Minimally Relevant)**:
- Very weak connection to the query
- May mention related terms but lacks substantial relevance
- Primarily off-topic with minimal connection
**0.0 (Not Relevant)**:
- No meaningful connection to the query
- Completely off-topic
- Contains no relevant keywords or concepts
## Analysis Principles:
1. **Semantic Understanding**: Look beyond exact keyword matches to understand the underlying meaning and context
2. **Query Intent**: Consider what the searcher is likely looking for based on the query
3. **Content Quality**: Factor in whether the post provides meaningful information related to the query
4. **Context Awareness**: Consider the context in which terms are used, not just their presence
5. **Specificity**: More specific, detailed content about the query topic scores higher than vague mentions
## Classification Rule:
- Posts with relevance_score at or above the classification threshold go into "relevant_posts"
- Posts with relevance_score below the threshold go into "irrelevant_posts"
- Every input post must appear in exactly one of the two arrays
## Output Rules:
- Return a JSON object with exactly two top-level keys: "relevant_posts" and "irrelevant_posts"
- Each entry in either array has exactly two fields: "post_id" (string, copied verbatim from the input) and "relevance_score" (float between 0.0 and 1.0)
- Do not invent extra fields (no "summary", no "reasoning", no "score", no "label")
- Do not rename the top-level keys
- Do not omit any input post
## Important Notes:
- Maintain consistency across similar posts in the batch
- Be objective and focus on content relevance rather than post quality, popularity, or personal opinions
- For borderline cases, err slightly toward inclusion rather than exclusion
## Required Output Format
Your response MUST be a single, valid JSON object conforming to this schema:
```json
{schema_json_string}
```User prompt
Please analyze the following batch of {content_type} for relevance to the search query. For each post, determine if it's relevant to the query and provide a relevance score from 0.0 to 1.0.
## Search Query:
"{query}"
## {content_type} to Analyze:
{posts}
## Instructions:
1. Read and understand the search query and its likely intent
2. Score each post for relevance to the query using the scoring guidelines (0.0-1.0)
3. Classify each post:
- Score ≥ {relevance_threshold} → put it in "relevant_posts"
- Score < {relevance_threshold} → put it in "irrelevant_posts"
4. Every input post must appear in exactly one array
## Required Output Shape:
- Top-level keys: "relevant_posts" and "irrelevant_posts" (both arrays)
- Each entry has fields:
- post_id: the exact id from the input
- relevance_score: float between 0.0 and 1.0
- No other fields. No other top-level keys.
## Output Format:
The required JSON output schema is provided in the system prompt.
GENERIC_RELEVANCE_SCORE_SYSTEM_PROMPT +
GENERIC_RELEVANCE_SCORE_USER_PROMPT
(36 calls in window)
System prompt
You are a relevance scorer. Your task is to evaluate how relevant each item is to a given context.
For each item, assign a relevance_score between 0.0 and 1.0:
- 1.0 = Highly relevant: directly and substantially addresses the context
- 0.7-0.9 = Relevant: clearly within scope, strong connection to the context
- 0.4-0.6 = Tangentially relevant: some connection but not a primary match
- 0.1-0.3 = Low relevance: weak or incidental connection
- 0.0 = Irrelevant: no meaningful connection
Scoring guidelines:
- Focus on topical and semantic alignment between each item and the context
- A high-quality item on an unrelated subject should still score low
- A brief item on a directly relevant subject should still score high
- Consider both explicit and implicit connections
Output rules:
- Return ALL items provided, one entry per item
- Each entry has exactly two fields: "id" (string, copied verbatim from the input) and "relevance_score" (float between 0.0 and 1.0)
- Wrap the entries in a single top-level key named "items"
- Do not invent extra fields (no "score", no "summary", no "reasoning", no "label")
- Do not rename the top-level key (it must be "items", not "analysis_batch", "results", "posts", etc.)
## Required Output Format
Your response MUST be a single, valid JSON object conforming to this schema:
```json
{schema_json_string}
```User prompt
## Scoring Context
{scoring_context}
---
## Items to Score
{items_list}
---
## Instructions
Score each item above for relevance to the scoring context.
Output exactly this shape and nothing more:
- Top-level key: "items" (an array, one entry per input item)
- Each entry has fields:
- id: the exact id from the input
- relevance_score: float between 0.0 and 1.0
Do not add any other fields. Do not rename the top-level key. Do not omit any item.
## Output Format
The required JSON output schema is provided in the system prompt.
JSON_REPAIR_SYSTEM +
JSON_REPAIR_USER
(9 calls in window)
System prompt
You are a JSON repair tool. The user gives you malformed or partial model output and a JSON Schema. Return ONLY a single valid JSON object that satisfies the schema, salvaging as much real content from the input as possible. Do not invent data for fields the input doesn't support — use the schema's allowed empty/null values. Output the JSON object only: no prose, no markdown, no code fences.
User prompt
JSON Schema:
{schema_json}
Malformed output to repair:
{raw_text}
Return only the corrected JSON object.