MCP App Store
by KeenEthics
TALIS Questionnaire Explorer icon

TALIS Questionnaire Explorer MCP Server for ChatGPT

By Leonard Najman

Overview

Ask about the TALIS (Teaching and Learning International Survey) questionnaires in plain language. The app answers from the published teacher and principal questionnaires (2008, 2013, 2018 and 2024). What you can do: - Find questions on a topic, by keyword or meaning. - Read the exact wording of a question, its items and its response scale. - Follow a concept across cycles and see which items were kept, added or dropped. - Compare two questionnaires, for example the 2018 and 2024 teacher questionnaires. - Find the dataset variable name of a question in each cycle. - See the routing of a question and the terms the questionnaire defines. Who it is for: researchers, analysts and survey designers who work with TALIS data and need the right item quickly. Limits: the app holds questionnaire content only, not survey responses. It can say what was asked, of whom and how, but not what respondents answered (no percentages, averages or country results). It is read-only and needs no account.

How to connect

  1. ChatGPT: Open the TALIS Questionnaire Explorer listing in the ChatGPT app directory and choose Connect. ChatGPT listing TALIS Questionnaire Explorer exposes 19 tools here.

Tools

talis_compare_instruments

ChatGPT
Compare two exact instrument printings exhaustively: what they share, what changed, and what only one of them asks. USE THIS FOR "what changed in the teacher questionnaire between 2018 and 2024?", "which questions were dropped or added?", "compare the 2024 Teacher and Principal questionnaires", and "what do these two instruments share?". Obtain exact keys from talis_list_instruments first. Order matters: put the earlier round on the left. Concepts are matched in three tiers, all exhaustive SQL: - same_concepts: the same question asked on both sides (concept keys are cross-cycle identity), each diffed for wording, items, answer options and routing; changed lists what differs. - curated_overlaps: different concepts joined by the curated cross-instrument map, e.g. a teacher question and its principal counterpart. - computed_candidates: suggestions only. Strong computed evidence (shared statements and a shared scale) between concepts no tier matched, e.g. a teacher and a principal question the curated map has not linked yet. Check one with talis_compare_questions; some strong pairs were left unlinked on purpose. - only_left / only_right: neither tier above. Between two rounds of the same questionnaire these are the dropped and the new questions. Between different respondents the curated map is conservative, so they are not proof that a topic is absent; caveat says which reading applies. For one source question, follow with talis_find_parallel_questions; for statement-by-statement change within one concept, talis_compare_items. Returns JSON: {found, left{instrument summary, domains[]}, right{...}, same_concept_count, changed_count, same_concepts[{question_concept_key, question_concept_label, left_question_keys[], right_question_keys[], changed[], unchanged}], curated_overlap_count, curated_overlaps[{cross_instrument_concept_key, cross_instrument_concept_label, left_concepts[], right_concepts[]}], only_left{count, concepts[]}, only_right{count, concepts[]}, computed_candidate_count, computed_candidates[{left_question_key, right_question_key, left_concept, right_concept, shared_item_count, stem_overlap, scale_match, evidence_score}], method, caveat}.

talis_compare_items

ChatGPT
Line a concept's individual items up across cycles: kept, added, dropped. USE THIS WHEN talis_get_concept reports that item_texts changed and you need to know WHICH items — the question actually being asked whenever someone wants to know whether a battery or scale is comparable over time. It is also the variable crosswalk: each row names the item's dataset variable in every cycle, so "what is the 2024 variable for tc3g22a?" is one call once talis_get_question(var_name=...) has given the concept. Items are matched on their words, ignoring case, punctuation and placeholder brackets, so a statement that survived unchanged lines up across cycles. missing_from therefore means "this statement is absent", which can mean dropped OR reworded. possible_rewordings pair statements that never appear in the same cycle and share many words. They are SUGGESTIONS scored by token overlap, not findings, and some will be different statements. Both wordings and their cycles are returned. Do not report a rewording as fact on the similarity score alone — read the two texts and judge. cycles narrows the comparison, e.g. [2018, 2024] for "which statements stayed the same between 2018 and 2024?". Concepts whose questions have no item text (where the item IS the question) return found=false; use talis_get_concept instead. cycles_asked is every cycle the concept appeared in; comparable_cycles is the subset with item-level text to align; cycles_without_items is the difference (single-item printings). stable_across_comparable_cycles and each row's in_all_comparable_cycles are scoped to comparable_cycles, NOT to the concept's full history — read them together with cycles_without_items. Returns JSON: {found, question_concept_key, cycles_requested[]?, cycles_asked[], comparable_cycles[], cycles_without_items[], item_count_by_cycle, distinct_items, stable_across_comparable_cycles, all_cycles_have_item_text, matrix[{item_text, present_in[], missing_from[], variables{}, in_all_comparable_cycles}], possible_rewordings[{similarity, possibly_reworded_from{years[],text}, into{years[],text}}], note}

talis_compare_questions

ChatGPT
Compare two exact questions side by side: wording, statements, answer options and routing. USE THIS FOR "compare the teacher and principal job satisfaction questions", "how does the 2018 version of this question differ from 2024?", and "do these two questions ask the same things?". It works for any pair: the same question in two rounds, a teacher question and its principal counterpart, or two questions that may overlap. Get the keys from talis_find_parallel_questions, talis_get_concept, talis_compare_instruments or talis_search_questions first. relationship says how the artifact links the pair: same_concept (one question in two printings, so differences are real changes), curated_link (different questions joined by the curated map), or unlinked (shared statements are overlap in wording, not proof they measure the same thing). Statements are matched on their words, ignoring case and punctuation. possible_rewordings pair one-sided statements that share most of their words, e.g. "...become a teacher" and "...become a principal". They are SUGGESTIONS: read both wordings before calling them the same statement. For a whole concept across every round, use talis_compare_items. Returns JSON: {found, relationship, left{question_key, year, instrument_key, instrument_name, respondent_type, question_number, question_concept_key, question_stem, ...}, right{...}, changed[], unchanged, shared_item_count, shared_items[{item_text, left_variables[], right_variables[]}], only_left_items[{item_text, variables[]}], only_right_items[], possible_rewordings[{similarity, left, right}], answer_options{left[], right[], same}, note}

talis_concept_coverage

ChatGPT
Which questions appear in which cycles. Structural, not a search. USE THIS FOR "which questions appear in all four cycles?", "what was dropped after 2018?", "what's new in 2024?", "how much of the 2013 questionnaire survived?". These arrive phrased like search requests, but routing them through semantic search returns a plausible subset instead of the correct complete answer — the worse failure. This tool is exhaustive. first_year=2024 lists what is new in 2024 and last_year=2018 what was dropped after 2018. cycles means present in ALL the listed cycles. min_cycles means present in at least that many. Combine with domain, instrument, or respondent_type to scope: years are counted inside the scope, so respondent_type='principal', first_year=2024 is what is new for principals. Without a scope, the exhaustive list intentionally contains separate Teacher, Principal and TKS identities. Returns JSON: {filters, count, concepts[{question_concept_key, concept_label, cycles, year_list[], first_year, last_year, records}]} Paged: limit (default 25) and offset. If has_more is true you are looking at a partial list — page before stating a total.

talis_find_cross_instrument_parallels

ChatGPT
Discover which questions in two instruments measure the same thing — both the curated links AND computed candidates the curated map misses. USE THIS WHEN comparing two questionnaires and asking "which questions appear in both?", "what does the Principal questionnaire ask that the Teacher one also asks?", "find the teacher counterpart batteries in the principal survey". It does in one call what would otherwise take dozens: for every question on each side it compares the item batteries, stems, response scales and stored embeddings, then returns the pairs with real overlap, ranked by strength of evidence. Two clearly separated sections: - curated_pairs: the artifact's own cross-instrument links. Safe to describe as measuring the same concept. - candidate_pairs: COMPUTED EVIDENCE, never asserted equivalence. shared_item_count is exact normalised-wording item matches (the strongest signal), stem_overlap is token Jaccard, scale_match is a shared response-scale concept, cosine_similarity is between stored embeddings, and evidence_strength (strong/moderate/weak) summarises them. Read both questions before calling any candidate equivalent — instruments adapt wording by respondent and can differ in reference period or response mode (e.g. one rates extent 1–4, the other Yes/No). DO NOT USE THIS to find one question's counterpart from a key you already have -> talis_find_parallel_questions. For the structural overview (counts, domain coverage, curated overlap only) -> talis_compare_instruments. A candidate absent here is not proof no parallel exists; it means it fell below the evidence thresholds. Raise limit to see weaker candidates. Returns JSON: {found, left{instrument_key, instrument_name, question_count}, right{...}, curated_pair_count, curated_pairs[{cross_instrument_concept_key, left_question_key, left_question_number, right_question_key, right_question_number, shared_item_count, stem_overlap, scale_match, cosine_similarity}], candidate_count, candidate_shown, candidate_pairs[{left_question_key, left_question_number, left_stem, right_question_key, right_question_number, right_stem, shared_item_count, left_item_count, right_item_count, shared_item_fraction, shared_items[], stem_overlap, scale_match, shared_scale_concepts[], cosine_similarity, evidence_strength, evidence_score}], method, caveat}

talis_find_parallel_questions

ChatGPT
Find questions parallel to one exact source question in another instrument. USE THIS FOR "is this Teacher question also asked of principals?" or "find the TKS counterpart of this question". The safe default, mode=curated, returns only questions sharing the artifact's cross-instrument concept key. Curated results are exhaustive and are never truncated by limit. mode=semantic or both is an explicit fallback for discovery. Those results are separately labeled cosine candidates: proximity does not prove identity, equivalence, shared measurement, or even the same construct. Read both original stems, then call talis_get_question on both keys to compare items, response scales, reference periods and routing. Never turn a semantic candidate into a finding without that evidence. Returns JSON: {found, mode, source, target_instrument, curated_available_count, curated_count, curated_results[], curated_note, semantic_candidate_count, semantic_candidates[{question_key, question_concept_key, cross_instrument_concept_key?, question_stem, cosine_similarity}], semantic_limit, semantic_caveat}. limit applies only to semantic candidates.

talis_get_concept

ChatGPT
Trace one question across TALIS cycles and report what changed. USE THIS FOR any cross-cycle question: "how did this change between 2018 and 2024?", "was this asked in 2013?", "compare the wording over time". question_concept_key is what makes two records in different cycles the same question. Text similarity is never identity — do not decide two questions match because they read alike; check they share a concept key. changes_vs_latest compares each earlier cycle to the most recent printing. Fields compared: question_stem, answer_type, question_instruction_text, item_count, item_texts, scale_options, plus routing_gated / routing_onward. Routing is compared by SHAPE only, because question renumbering makes routing strings look changed when nothing changed. form_coverage is reported per cycle but NOT diffed — it exists only for 2024, so diffing it flags 91% of all comparisons. Case and punctuation differ between cycles faithfully to the printed questionnaires ('Strongly Disagree' in 2008, 'Strongly disagree' in 2024). That is fidelity, not error. Quote the original. Returns JSON: {found, question_concept_key, label, respondent_scope, cycles[], record_count, domains[], domain_note?, timeline[], latest_year, changes_vs_latest[{year, question_key, changed[], unchanged, form_coverage}]}

talis_get_question

ChatGPT
Retrieve everything about one printed question. Use this before describing any question's wording, items, scale or variables. Accepts either a question_key ('2024:teacher_questionnaire:q48') or a var_name ('tt4g48a'), not both. A key without its ':p1' part suffix returns every part. IMPORTANT — printed grain is not variable grain. One printed question can be several records, one record several rows, one row several cells. When count > 1 read note: part_type 'alternative' means ONE question printed twice for two populations (2024 Q48), and must never be reported as two questions. Text may contain placeholders like <target class> or [<ISCED 2011 Level X> / 15-year-old]. These are country-substitution markers and are part of the content — reproduce them verbatim, do not resolve or drop them. For grid items, item_text is 'row wording – column label' and row_wording gives the row on its own. Returns JSON: {found, count, records[], note?}. Each record has the full v_question row plus items[] (with var_name, row_label, column_label, item_text, row_wording, answer_type, scale_name), variables[], scales[] (each with scale_key, scale_concept_key and all options), routing_in[], routing_out[], definitions[], row_layout[], concept_cycles[].

talis_get_scale

ChatGPT
Get one response scale, its options in each cycle, and the questions that use it. USE THIS FOR "which questions use a four-point agree scale?", "did the response options change?". Find a scale_concept_key on the output of talis_get_question. Returns JSON: {found, scale_concept_key, scale_concept_label, by_cycle[{year, scale_key, scale_name, options[{code, label, position}]}], used_by[{year, question_key, question_number, question_stem, question_concept_key}], note}

talis_list_definitions

ChatGPT
List the terms TALIS defines for respondents, and how they changed. USE THIS FOR "how does TALIS define professional development?", "what counts as mentoring?", and above all BEFORE BUILDING ANY TREND on a concept that depends on a defined term. Definition drift is a silent threat to comparability. Placements can attach to a section, question or item, so the server resolves each one back to its instrument before comparing texts. Summary rows are separated by respondent type; instrument filters may select one exact printing or a same-named family across cycles. Omit term to list all defined terms. Supply one to also get the full text of each version with the cycle and question it was attached to. Returns JSON: {filters, count, terms[{definition_term, respondent_type, instruments[], cycles, year_list[], distinct_texts, placements, text_changed_across_cycles}], placements[{definition_term, year, definition_text, display_mode, placement_type, question_key, section_key, item_key, instrument_key, instrument_name, respondent_type}], note}

talis_list_domains

ChatGPT
List the respondent-scoped TALIS domains with keys and concept counts. USE THIS FIRST when you do not yet know what TALIS covers, or when you need valid domains values for talis_search_questions. Use instrument or respondent_type whenever the question is about one questionnaire; the unfiltered catalogue contains Teacher, Principal and TKS domains. Domains are scoped identities, even when their labels resemble one another. Returns JSON: {filters, count, domains[{domain_concept_key, domain_label, respondent_scope, concept_count, first_year, last_year}], note}

talis_list_instruments

ChatGPT
List every questionnaire/instrument printing with its exact stable key. USE THIS FIRST for "which questionnaires are available?", before comparing instruments, or whenever a user says only "the Teacher questionnaire" and the cycle is unclear. Instrument names repeat across cycles, so downstream cross-instrument tools require the exact instrument_key returned here. Optional year and respondent filters narrow the inventory without blending Teacher, Principal and Teacher Knowledge Survey counts. Counts are exhaustive SQL summaries, not retrieval estimates: questions, concepts, items, variables, sections, conditional questions, domains, scales and definitions. question_count counts printed question records, so a question printed in two alternative versions (2024 teacher Q48) counts twice; numbered_question_count counts question numbers, which is the figure to quote for questionnaire length. Returns JSON: {filters, total, count, offset, has_more, next_offset, instruments[{year, instrument_key, instrument_name, respondent_type, question_count, numbered_question_count, concept_count, item_count, distinct_variable_count, section_count, conditional_question_count, domain_count, scale_count, definition_count}]}. Paged with limit and offset; page before claiming the inventory is complete.

talis_list_scales

ChatGPT
Browse and filter the response scales, by shape rather than by name. USE THIS FOR "which questions use a four-point agree scale?", "are there any frequency scales?", "did any response scale change between cycles?" (changed_only=true lists just those). Use instrument/respondent filters before reporting questionnaire-specific counts; the unfiltered catalogue spans all instruments. Grouped by scale_concept_key, which is cross-cycle identity for scales. variants counts per-cycle printings and does NOT imply the options changed — options_changed_across_cycles is the field that tells you that, and it compares option labels normalised for case and punctuation. 2008 prints "Strongly Disagree" and 2024 "Strongly disagree"; that is fidelity to the printed questionnaires, not a change. Display latest_shape, which is the original wording. Paged: limit (default 25) and offset. Returns JSON: {filters, total, count, offset, has_more, next_offset, scales[{scale_concept_key, scale_concept_label, variants, cycles[], min_options, max_options, latest_shape, distinct_shapes, options_changed_across_cycles}], note}

talis_list_structure

ChatGPT
The table of contents for a questionnaire, in printed order. USE THIS FOR "show me the 2024 Teacher questionnaire", "how many questions are in the 2013 Principal questionnaire?", or "what is in this instrument's school-climate section?". ALWAYS QUALIFY A YEAR BY INSTRUMENT when the user means one questionnaire. Every cycle now contains multiple instruments; 2013 has Teacher, Principal and the Mathematics Teacher Module, and 2024 has Teacher, Principal and TKS. A year-only call intentionally returns all of them. The available field lists every exact year/instrument pair. Returns JSON: {filters, count, questions[{year, instrument_key, instrument_name, section_position, section_heading, question_key, question_number, question_part, part_type, question_stem, answer_type, form_coverage, question_concept_key, question_concept_label, domain_concept_key, domain_concept_label}], available[]} Paged: limit (default 25) and offset. If has_more is true you are looking at a partial list — page before stating a total.

talis_list_topic

ChatGPT
Exhaustively list the questions in a broad topic area. USE THIS WHEN completeness matters: "list all the questions on professional development", "what does TALIS ask about school climate?", "everything on teacher appraisal". It works by resolving the topic to ONE respondent-scoped domain and then enumerating that domain in full, so the answer is complete by construction. Similarity ranking is not: on "professional development" it returned 0 of the 6 relevant concepts, while enumeration returned 6 of 6. DO NOT USE THIS WHEN: - the user wants the questions most relevant to a narrow topic rather than everything in an area -> talis_search_questions. A domain can hold many concepts, and returning all of them for a precise question buries the answer. - the user is quoting item wording -> talis_search_items. - the topic is not a broad area at all (a single variable, a specific question number) -> talis_get_question. Domains are cycle-scoped, and some concepts sit in more than one because they were re-filed between cycles. first_year/last_year on each concept are the years it sat in THIS domain, not the years it was asked — use talis_get_concept for the latter. CHECK caveats BEFORE ANSWERING. It tells you when the topic did not map cleanly onto a domain, or when a second domain scored nearly as high. Pass that on rather than presenting a truncated list as complete. Scope domain resolution with instrument or respondent_type whenever the request is about one questionnaire. Returns JSON: {query_type, method, resolved{domain_concept_key, domain_label, score, covers (only when unfiltered)}, alternatives[], count, concepts[{question_concept_key, question_concept_label, first_year_in_domain, last_year_in_domain, cycles_asked[], record_count}], caveats}. cycles_asked is every round the concept was asked in; the _in_domain years are only its span in this domain and filter. Paged: limit (default 25) and offset. If has_more is true you are looking at a partial list — page before stating a total.

talis_list_variables

ChatGPT
Map dataset variable names to the questions they came from, in bulk. USE THIS WHEN someone is working with TALIS microdata and needs to know what their columns asked: "what are all the 2024 variables?", "which variables belong to the well-being question?", "which variables cover professional learning?" (domain), "what does the tt4g65 family cover?". The artifact maps variables across instruments; filter by instrument before describing one microdata file. talis_get_question(var_name=...) answers one variable at a time and gives far more detail per variable; this is the bulk view for orienting in a dataset. Note one row can carry several variables — a grid row asked twice yields 'btg11a1, btg11a2' — so variable counts exceed item counts. Paged: limit (default 25) and offset. Filter before paging: an unfiltered call spans all cycles and all instruments. Returns JSON: {filters, total, count, offset, has_more, next_offset, variables[{var_name, year, instrument_name, question_key, question_number, question_part, question_concept_key, question_concept_label, row_label, column_label, item_text, answer_type, scale_name}]}

talis_search_items

ChatGPT
Find a specific sub-item (a single row of a grid) by its wording. USE THIS WHEN the user is quoting or half-remembering the actual text of one item: "which question has 'my job negatively impacts my mental health'?", "there's a statement about students getting on well with teachers". This is lexical (BM25) matching, and it beats semantic search on this job precisely because the user is reproducing real wording. DO NOT USE THIS for topics or paraphrases -> talis_search_questions. Scores are BM25: corpus-relative, unbounded, and NOT comparable to the cosine scores from talis_search_questions. Use them to rank within this result set only. There is no absolute "good score" threshold; judge a hit by reading its passage. Each hit carries variables, the dataset variable names for that row, and question_key so you can call talis_get_question for full context. Scope by instrument/respondent when the remembered wording is known to come from one questionnaire; global BM25 statistics span every instrument. Returns JSON: {query_type, method, score_scale, filters, count, hits[{target_key, question_key, question_concept_key, question_concept_label, domain_concept_label, year, instrument_name, question_number, question_stem, variables[], passage, score}]}

talis_search_questions

ChatGPT
Find TALIS questions about a topic, by meaning rather than by keyword. USE THIS WHEN the user names a subject and wants the questions that cover it: "is there anything on teacher burnout?", "questions about how principals give feedback". It matches on meaning, so 'teacher burnout' finds 'Well-being and stress' despite sharing no words. DO NOT USE THIS WHEN: - the user wants EVERY question in an area ("list all the questions on professional development") -> talis_list_topic. One query vector cannot sit close to all six concepts in a domain; this tool will return a plausible subset, which is a worse failure than no answer. - the user is quoting remembered item wording -> talis_search_items. - the user already has a key or variable name -> talis_get_question. Results are grouped by question_concept_key, which is cross-cycle identity: one hit means one question, however many cycles it appeared in. instrument and respondent_type are correctness filters, not cosmetic: use them whenever the request names one questionnaire or respondent. IF hits IS EMPTY, nothing passed the similarity cutoff: say there is no strong match. closest then lists up to three nearer concepts below the cutoff. Present them as the closest available, clearly weaker than a match, and check one with talis_get_question before relying on it. If closest is empty too, nothing is close: say no question was found, noting that search cannot prove a topic is absent. Never lower the cutoff or present a closest concept as a finding. Returns JSON: {query_type, method, cutoff, filters, count, hits[], near_miss_floor, closest[] (same shape as hits; only when hits is empty), note} where each hit is {question_concept_key, question_concept_label, score, cycles[], best_question_key, best_question_stem, best_year, instrument_name, matched_passage, domain_concept_key, domain_concept_label, best_target_key, best_grain, also_asked_as[]}. matched_passage is the text that actually matched, so you can show why. also_asked_as holds the same question family in other instruments (e.g. the Teacher Knowledge Survey or principal version), folded in through the curated cross-instrument map, each with its own score.

talis_trace_routing

ChatGPT
Show which respondents reached a question, and what it opens up next. USE THIS BEFORE INTERPRETING ANY RESPONSE COUNTS. Some questions in each cycle are conditional. If a question is gated, the people who did not answer it were never asked — that is a different thing from missing data, and treating it as missing will misstate a denominator. who_was_asked states the condition in plain terms, naming the question it depends on and the questionnaire forms the question was printed on. A question is gated by a filter that targets it (gate_kind 'filter') or by a skip that jumps over it ('skipped_over'). The destination of a skip is NOT gated: reached_by_skip lists the skips that land there early, and everyone else reaches it in order. same_condition_also_gates lists the sibling questions governed by the same condition, because one condition commonly gates a run of questions (in 2024 the condition on Q4 gates Q5, Q6 and Q7 together). Returns JSON: {found, question_key, year, instrument_name, question_number, question_stem, form_coverage, is_conditional, who_was_asked, gated_by[{gate_kind, routing_type, trigger_code, routing_condition, source_question_key, source_question_number, source_item_key, skip_destination_key}], reached_by_skip[{routing_type, routing_condition, source_question_key, source_question_number}], opens[{routing_type, routing_condition, target_question_key, target_question_number}], same_condition_also_gates[{routing_condition, target_question_key, question_number, question_stem}]}

Capabilities

Search Questions By Topic Or WordingShow Exact Question Wording And Response ScalesCompare Items Across Survey CyclesCompare Two QuestionnairesFind Dataset Variable Names Across CyclesTrace Question Routing And Defined Terms

Example Prompts

Click any prompt to copy it.

Similar Apps

Listed on MCP App Store

Listed on MCP App Store

Add this badge to your site or README. It links to this page.

HTML

<a href="https://mcp-app-store.keenethics-labs.com/app/talis-questionnaire-explorer"><img src="https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg" alt="Listed on MCP App Store" width="220" height="44"></a>

Markdown

[![Listed on MCP App Store](https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg)](https://mcp-app-store.keenethics-labs.com/app/talis-questionnaire-explorer)