IA-QA Eval & RAG — 38 tools

Query the IA-QA methodology knowledge base. Returns structured testing guidelines, assertion strategies, thresholds, best practices, and relevant MCP tools for a given topic. Call without a topic to list all available topics. Topics: llm-unit-testing, rag-pipeline, prompt-stability, prompt-ab-testing, embedding-quality, eval-framework, semantic-testing, auto-testing, security, api-testing, ci-cd, multimodal, llm-data-security, agent-observability, pro-tips, learning-paths, golden-dataset, selector-drift, qa-recipes, playbooks. Not sure where to start testing an LLM, RAG pipeline or agent? Call without a topic (or with "start-here"): it maps what you are testing to the tools to call and the output field to gate CI on. A plain question such as "how do I test my RAG" also resolves to the right topic.

Search available MCP tools by keyword or category before calling them. Returns matching tool names, descriptions, and optionally their inputSchemas. Call this when you are unsure which tool to use or want to explore the catalogue. Categories: data, encoding, text, llm, qa, rag, dev, security, web.

Lexical hallucination check: verifies an LLM answer's words, numbers and polarity against the provided source/context. Fast, deterministic, no API key needed. Each answer sentence is aligned to its best-matching source sentence, so a number only counts as support when it sits on the SAME statement ("founded in 1998" is not grounded by "sold 1998 units"), and a negation or antonym flip against that sentence returns verdict "contradicted" — the corrupted-fact hallucination that reuses source vocabulary. Limitations: still lexical — it cannot follow a paraphrase, a synonym, or multi-sentence reasoning, so a "well_grounded" verdict means "nothing lexical found", never "verified true". For entailment use run_semantic_tests (NLI/embedding) or a calibrated judge.

Parse a .ia-eval.yaml LLM test suite, call the specified LLM model for each scenario, run all configured scorers, and return a structured JSON report with per-scenario Pass/Fail verdicts and a Markdown summary. Use list_local_tests to discover available test files.

Semantic assertion primitive: compare actual vs expected text pairs using cosine similarity + ROUGE-L. Two modes: tfidf (default, free, no API key) or embeddings (OpenAI text-embedding-3-small, BYOK, true semantic similarity). Returns per-case PASS/FAIL verdicts and an overall verdict. CI-ready: pipe the JSON verdict field to gate a build.

Generate a ready-to-commit GitHub Actions workflow that gates a build on IA-QA. Two gate types, combinable: "eval_contract" runs a .ia-eval.yaml through ia-qa-com/eval-action@v1 (LLM quality gate, needs a provider API key as a repo secret), and "cli_checks" runs deterministic primitives via npx @ia-qa/cli (secret scan, prompt-injection scan, security headers…) whose exit code fails the build. Deterministic template — no LLM call, no API key, same inputs give the same file. Returns the YAML, the secrets to create, and the remaining steps. Pair with generate_eval_yaml to produce the contract itself.

Generate a "needle in a haystack" test: embeds a target fact into a large block of filler text at a specified position. Use this to test LLM context window retrieval accuracy. Returns the full haystack, the question to ask, and metadata. No API key needed.

Generate a complete .ia-eval.yaml evaluation contract from a plain-language description of what your LLM should do. Uses Groq openai/gpt-oss-20b (server-side, no API key needed). Returns ready-to-run YAML for the LLM Test Runner (run_eval_contract). Picks appropriate evaluators (cosine_similarity, contains_check, hallucination_check, etc.) based on the task type.

Reference-free stability primitive: instead of comparing an answer to a ground truth, it checks that an assistant's answer stays INVARIANT when the QUESTION is transformed (typo, casing, paraphrase, reordering, translation). Catches the failure class no reference answer can expose — an assistant that handles one phrasing well and a trivial variant of it badly. You bring the outputs (no model is called), so it is deterministic and free in tfidf mode. Relations: case (θ .95), typo (.90), paraphrase (.80), reorder (.80), translation (.75, embeddings only), specialization (.60, ADVISORY — directional, never gated). Returns PASS / FAIL / INVALID, where INVALID means the BASE answer was a refusal or too short so invariance was never measurable — an assistant that refuses every variant would otherwise score a perfect 1.0. Use run_semantic_tests alongside it: invariance without a correctness floor is a green light for a broken assistant.

Test a SKILL.md in the open Agent Skills format (https://agentskills.io/specification — read by Claude Code, Codex, GitHub Copilot, Cursor, Gemini CLI, OpenCode and others) in two stages. (1) lint: the spec's frontmatter rules checked without a model (name 1-64 chars, lowercase/digits/single hyphens, matches dir_name when given; description 1-1024 chars; compatibility ≤500; metadata string map) — any error returns BLOCK and calls no model; gate on lint.valid, it is deterministic. (2) triggering, measured the way an agent decides: the model sees ONLY the name + description of this skill and of the other installed skills (other_skills, or six built-in neighbours) and answers which one it would load, at temperature 0. trigger_accuracy = share of requests routed correctly (should load / should not); verdict PASS ≥80, FIX ≥50, else BLOCK. Pass your own scenarios [{prompt, should_trigger}] for a reproducible gate; without them requests are generated (half should-load, half near misses) and the result is a signal, flagged by scenarios_source. step_adherence is keyword overlap on the skill body's run and never enters the verdict. A request the model never answered readably (rate limit, empty output) is not counted as a miss: the run is INCONCLUSIVE instead. Uses the server's default Groq model (server key, rate-limited). Pass api_key + model to use your own provider; the SKILL.md and scenarios are sent to ia-qa.com and to that provider.

Validate an LLM response against QA criteria: format checks (JSON, code, markdown), content rules (must-include, must-not-include), length constraints, language detection, and safety patterns. Essential for QA testing LLM-powered features.

Compare two ALREADY-PRODUCED outputs (e.g. model A vs model B on the same task) side by side. Returns deterministic metrics (token cosine, ROUGE-L, Jaccard, length/structure deltas, JSON diff) and a verdict. If a `reference` (ground truth) is given, scores each output against it and picks the closer one. If `model` + `api_key` are given, an LLM judge also picks a qualitative winner for the task. No re-execution — you bring the outputs.

Semantically analyze N already-produced model outputs for the SAME task (the MCP counterpart to the LLM Sandbox). Without a reference: computes consensus — pairwise cosine agreement, the most-representative output, and the outlier. With a `reference` (ground truth): also ranks every output by closeness (token cosine + ROUGE-L composite) and names the closest. No LLM, no key, runs locally. The reference mode is gate-able in CI; the consensus/outlier mode is a signal (no reference to fail against). You bring the outputs (2+). For a 2-way head-to-head with structural JSON diff use compare_responses instead.

Define a test suite for a prompt: provide the system prompt, user prompt, and expected output criteria. Returns a test plan with scored rubric — use this as input for manual or automated LLM evaluation.

Validate that an LLM output matches an expected format: JSON, Markdown, code block, bullet list, numbered list, table, YAML, XML, or custom regex. Essential for structured output testing.

Compare multiple LLM responses to the same prompt and detect inconsistencies using Jaccard word-overlap similarity and fact drift (number comparison). Fast, deterministic, no API key needed. Limitations: relies on surface-level word matching — "Paris is the capital of France" vs "Paris is the French capital" may score low despite semantic equivalence. For true semantic consistency, use run_semantic_tests with embedding mode. Essential for determinism testing.

Validate that an LLM JSON output matches a JSON Schema definition. Tests required fields, types, enums, nested objects, and arrays. Critical for function-calling and structured output testing.

Score an LLM response against the criteria you pass: coverage of expected_keywords and compliance with max_length. Returns a 0-100 score over the criteria actually measured — or total_score null with the reason when none is given, because relevance and correctness cannot be read off word overlap (a correct one-word answer shares no words with its question). For meaning against a reference answer, use run_semantic_tests. Also returns unscored signals: question-term overlap, average sentence length, markdown.

Analyze a multi-turn conversation for context retention, topic drift, instruction following, and repetition. Accepts messages array [{role, content}]. Essential for chatbot QA.

Analyze a Chain-of-Thought (CoT) or reasoning trace from an LLM. Detects step count, logical flow, conclusion presence, backtracking, and estimates reasoning depth. Useful for o1/o3/DeepSeek-R1 evaluation.

Generate an A/B test report comparing two prompts or model configurations. Accepts arrays of scores and returns mean, median, std deviation, improvement percentage, and a Welch t-test at 95 %: `winner` names a variant only when the difference is significant — otherwise "no_significant_difference", or "insufficient_samples" under 2 scores per variant.

Compute text similarity between reference and hypothesis using multiple metrics: Cosine (BoW, TF-IDF), Jaccard, ROUGE-1, ROUGE-2, ROUGE-L, and BLEU. No API key needed. Ideal for LLM eval (expected vs actual), RAG quality checks, and NLG benchmarking. Supports batch mode.

Run a test suite against a Vision-Language Model (VLM) — send an image (URL or base64) + N test cases (each with a question + assertion) to GPT-4o, Claude 3.5, or Gemini. Returns per-case PASS/FAIL verdicts, a pass rate, an overall PASS/WARNING/FAIL verdict (customizable threshold), and latency stats. Assertion types: contains, not_contains, json_format, min_length, max_length, semantic_contains (TF-IDF cosine similarity ≥ 0.4). BYOK: requires your own API key for the target provider.

Compute text similarity using local algorithms (Bag of Words, TF-IDF, Character N-grams). No API key needed — runs entirely in-process. NOT real embeddings: for true semantic similarity with vector embeddings, use run_semantic_tests with mode="embeddings" and your OpenAI API key. Supports single pair or batch mode with pipe-separated pairs. Useful for RAG retrieval testing, semantic search evaluation, and text deduplication.

Rank an array of text chunks by relevance to a query using TF-IDF scoring. Simulates retrieval ranking for RAG testing without needing embeddings or an API.

Compute BM25 relevance score between a query and one or more documents. BM25 is the industry-standard keyword-based ranking algorithm used in Elasticsearch, OpenSearch, and Weaviate hybrid search. Returns ranked results with normalized scores.

Assemble a complete RAG (Retrieval-Augmented Generation) prompt from retrieved context chunks and a user query. Handles token budgeting, citation numbering, system instruction injection, and source attribution.

Evaluate RAG retrieval quality: rank passages against a query and compute Precision@k / Recall@k plus a PASS/FAIL CI verdict from ground-truth relevance labels. Three modes, all keyless except the last. (1) BYO scores — give each passage the `score` your own reranker produced (Cohere, Jina, a self-hosted NIM, a cross-encoder): deterministic, offline, and it evaluates YOUR reranker rather than someone else's. This is the mode to gate CI on. (2) Default, no scores and no key — ranks with local BM25, a lexical keyword baseline: it answers "does a keyword floor already surface my relevant passages?", never "is my neural reranker good". (3) Live NVIDIA reranker — supply `api_key` for an NVIDIA account that still has reranking entitlement; NVIDIA retired its hosted reranking endpoints on 2026-05-18, so this one is for accounts that were grandfathered in.

Estimate the token count of a text string using the cl100k_base approximation (~4 chars/token). Call this BEFORE sending any text to an LLM API to check if it fits within the model context window and to estimate cost. Returns token estimate, character count, and word count.

Extract the first valid JSON object or array embedded in chaotic LLM output (surrounded by markdown fences, prose, or explanatory text). Handles ```json blocks and inline JSON. Call this whenever an LLM returns structured data mixed with explanation text instead of raw JSON.

Estimate the API cost in USD for a given model and token counts. Supports all major 2024–2026 models: GPT-4o, GPT-4.1, o3, o4-mini, Claude Opus 4, Claude Sonnet 4/4.5, Gemini 2.5 Pro/Flash, DeepSeek V3/R1, Grok 3, and legacy models.

Plan token allocation across system prompt, user input, context/RAG chunks, and expected output. Warns if budget exceeds model context window. Supports 25+ models.

Given an array of message objects [{role, content}], estimate total token usage and check if it fits in the target model's context window. Warns about truncation risk.

Get detailed specs for an AI model: context window, pricing per 1K tokens, knowledge cutoff, provider, multimodal support, reasoning capabilities, and feature list. Covers 30+ models from OpenAI, Anthropic, Google, DeepSeek, Meta, Mistral, Cohere, xAI.

Compare 2-5 AI models side by side: context window, pricing, multimodal, reasoning capabilities, and provider. Returns a comparison table with a recommendation based on your use case.

List all LLM models available on ia-qa.com with their provider, API endpoint, and capabilities. Filter by provider name (e.g. "Groq", "HuggingFace", "OpenAI") or return the full catalog. Use this to discover which models are available before calling an LLM API, or to compare providers.

Generate text with an open-source model hosted on Groq, on the server's key — no API key needed. Models: openai/gpt-oss-20b, openai/gpt-oss-120b (default "openai/gpt-oss-20b"). HuggingFace models are not served on the server key. A reasoning model that spends max_tokens thinking is retried once at 4096. Rate-limited to prevent abuse.

Give honest usage feedback on an IA-QA MCP tool. Provide a score (1-5) and a comment. Rate low (1-2) if the tool was wrong, irrelevant, or a poor fit; rate high (4-5) only if it genuinely solved your need. Ratings are read by the IA-QA maintainers and are not published, one vote per agent and tool a month. Comments are screened before storing: one carrying a credential is not stored and personal data is replaced by placeholders — describe what happened rather than what a model should do. Skip rating routine successes — we want signal, not praise. Example: rate_tool({ tool_name: "format_json", score: 2, comment: "Tried to pretty-print a JSON5 file, it rejected trailing commas — not usable for my case." })