TriQuest: QA Retrieval & Curriculum Design

• TriQuest is a term applied to two distinct systems: one for multimodal QA retrieval using a tri-encoder architecture, and another for AI-assisted curriculum design.
• The tri-encoder model uses separate transformers for questions, text passages, and tables, leveraging dot product scoring and contrastive learning for effective retrieval.
• The AI-copilot platform integrates LLMs, knowledge graphs, and a guided GUI workflow to streamline interdisciplinary lesson-plan design and evaluation.

Searching arXiv for papers using the name "TriQuest" and related work on tri-encoder multimodal retrieval. TriQuest is a name applied to two distinct systems in the arXiv record. In the 2021 retrieval paper by Kostić et al., TriQuest denotes a tri-encoder architecture for open-domain extractive question answering that jointly embeds questions, free-form text passages, and tables into a shared vector space, enabling first-stage retrieval over heterogeneous evidence (Kostić et al., 2021). In the 2025 curriculum-design paper, TriQuest denotes an AI-copilot platform that combines LLMs, knowledge graphs, vector retrieval, and a GUI-guided workflow to support interdisciplinary lesson-plan design and review (Wang et al., 3 Oct 2025). The shared label therefore spans two separate technical contexts: multi-modal information retrieval and human–AI collaborative educational design.

1. Terminological scope

The term TriQuest is not attached to a single canonical system. In the cited literature, it designates two unrelated artifacts with different objectives, data modalities, and evaluation protocols.

Usage of “TriQuest”	Domain	Core formulation
TriQuest in 2021	Multi-modal QA retrieval	Three encoders $(E_q, E_t, E_{\rm tab})$ in a shared embedding space
TriQuest in 2025	Interdisciplinary curriculum design	Three-layer platform with OSS, MySQL, Neo4j, LLM modules, and GUI workflow

This dual usage matters because the two systems share neither an evaluation benchmark nor a task definition. One addresses retrieval of answer-bearing texts and tables for extractive QA; the other addresses teacher-facing lesson-plan generation, review, and evaluation in interdisciplinary education (Kostić et al., 2021, Wang et al., 3 Oct 2025).

2. TriQuest as a tri-encoder retriever

In the retrieval setting, TriQuest uses three separate transformers to encode three distinct input types—questions, free-form text passages, and tables—into a single shared embedding space. The architecture is explicitly factorized into a question encoder $E_q$, a text encoder $E_t$, and a table encoder $E_{\rm tab}$, producing $d$-dimensional vectors $h_q$, $h_t$, and $h_{\rm tab}$, respectively. In the reported experiments, each encoder is instantiated with either BERT-small, specified as 4 layers, hidden size 512, 8 heads, and approximately 29 M parameters, or TaPas-small, described as the same size and pre-trained on tables (Kostić et al., 2021).

Input representations are modality-specific. Questions and text passages use standard WordPiece tokenization with [CLS] prepended and [SEP] appended. Tables are flattened into a single sequence consisting of [CLS], page title, section title, table caption, each row concatenated cell-by-cell, and [SEP]. When a TaPas encoder is used, special row and column embeddings assign each token a 2D table position. Embeddings are taken from the final hidden state of the [CLS] token:

$$h_q = E_q([CLS]\ \text{question}\ [SEP]) \in \mathbb{R}^d,$$

$$h_t = E_t([CLS]\ \text{text}\ [SEP]) \in \mathbb{R}^d,$$

$E_q$0

No extra projection layer was used in the final models. The authors report that they experimented with an MLP projection but found no significant gains. Within the paper’s own interpretation, the main advantage over a bi-encoder is modality-specific encoding: separate transformers can specialize in linguistic structure versus tabular structure while still mapping all outputs into the same $E_q$1 space.

3. Scoring, retrieval, and optimization

At inference time, TriQuest assumes precomputed embeddings for all text passages $E_q$2 and all tables $E_q$3. A new question $E_q$4 is encoded as $E_q$5, and similarity is computed against both inventories. The scoring function used in the final implementation is the dot product,

$E_q$6

where $E_q$7 is either $E_q$8 or $E_q$9. The paper notes that cosine similarity can also be used,

$E_t$0

but reports that dot product worked best (Kostić et al., 2021).

The retrieval pipeline is straightforward: encode the question, score it against every text embedding and every table embedding, merge the two scored lists, sort by descending score, and return the top-$E_t$1 documents, which may be texts or tables. This is a first-stage retriever and does not perform cross-modal attention or any interaction between question and document at encoding time.

Training uses a contrastive, in-batch-negative-aware softmax loss à la DPR. Each batch contains $E_t$2 questions. For each question $E_t$3, the training signal consists of one positive document $E_t$4, which may be either a gold text passage or a gold table, and one hard negative $E_t$5 obtained by running BM25 and selecting the top result that does not contain the answer string. The other $E_t$6 questions’ positives and negatives are treated as additional in-batch negatives. With

$E_t$7

and

$E_t$8

the per-example loss is

$E_t$9

and the batch loss is

$E_{\rm tab}$0

Negative mining is performed with BM25 over a 21 M-passage index plus a 7 M-table index. Optimization uses Adam with learning rate $E_{\rm tab}$1, linear warmup over 10% of total steps, linear decay, 10 epochs, and batch size 28 for the tri-encoder versus 38 for the bi-encoder, with the difference attributed to GPU RAM constraints.

4. Dataset construction and empirical profile

The retrieval paper introduces the MultiModalRetrieval dataset by merging four existing sources: Natural Questions for questions whose answers lie in text, and NQ-Tables, WikiSQL filtered to context-independent questions, and OTT-QA for questions answered by tables. The resulting train/test split contains 120 239 training examples and 4 937 test examples. Hard negatives for training are mined using the same BM25-based procedure used during model construction (Kostić et al., 2021).

Evaluation is performed over a search space of 500 000 random passages, including the gold ones, together with 656 166 tables, for approximately 1.15 M documents. The principal metric is recall@$E_{\rm tab}$2 for $E_{\rm tab}$3. BM25 with two separate indices, one for text and one for tables, serves as the sparse baseline. The paper also reports recall on each uni-modal dataset using 1000 random test questions each.

The reported results are modality-sensitive. On NQ, BM25 achieved 53.3 / 59.8 recall@10 / recall@20, the bi-encoder 70.1 / 76.0, and the tri-encoder 69.1 / 75.0. On OTT-QA, the corresponding numbers were 40.2 / 45.6, 72.9 / 78.0, and 73.8 / 79.7. On NQ-Tables they were 56.6 / 65.1, 84.9 / 91.2, and 86.4 / 91.6. On the combined MultiModalRetrieval benchmark they were 50.7 / 57.0, 55.2 / 61.8, and 56.1 / 62.3. The WikiSQL variants behaved differently: on WikiSQL, BM25 obtained 42.1 / 47.1 while the tri-encoder reached 20.2 / 26.8; on WikiSQL_ctx-ind, BM25 obtained 61.2 / 67.2 while the tri-encoder reached 30.8 / 38.2.

The paper’s own summary is that dense transformer embeddings outperform sparse embeddings on four of six evaluation datasets, and that tri-encoders improve retrieval performance over bi-encoders when both text and table evidence must be handled jointly. At the same time, BM25 remains strongest on very high lexical-overlap datasets such as WikiSQL, and dense models lag when overlap exceeds 50%. This suggests that TriQuest’s advantage is strongest in low-overlap tabular or genuinely multi-modal settings rather than in purely lexical retrieval regimes.

5. TriQuest as an AI-copilot platform for curriculum design

In the 2025 educational setting, TriQuest is described as a three-layer architecture. The data layer stores raw curriculum standards, textbook editions, exemplary lesson-plan cases, and evaluation criteria in an Object Storage Service, deposits extracted knowledge triples in a MySQL relational database, and then imports them into Neo4j to form three distinct interdisciplinary knowledge graphs: Content KG, Normative KG, and Evaluation KG. The business layer contains five core modules: knowledge-graph construction, enhanced retrieval prompt engineering, LLM-guided human–computer collaborative lesson-plan generation, automated lesson-plan analysis and evaluation, and intelligent Q&A using the interdisciplinary design knowledge graph. The application layer exposes four user-facing modules: cross-disciplinary lesson-plan learning, lesson-plan design, knowledge-graph visualization, and evaluation and feedback (Wang et al., 3 Oct 2025).

LLMs, explicitly including iFlytek Spark and Qwen-1.5-1.8b-chat, are invoked via API in multiple modules. Vector databases and similarity search are used for case retrieval and KG-based prompt augmentation. The interface is organized as a step-by-step guided workflow that mirrors the platform’s view of instructional design practice:

1. Theme & Case Selection
2. Learner Analysis
3. Curriculum-standard alignment
4. Instructional-content assembly
5. Learning objectives formulation
6. Assessment design
7. Activity design & rationale
8. Theoretical foundation
9. Tools & resources selection

At each stage, teachers can select built-in prompt templates, invoke AI to draft text, and manually refine the result. The GUI displays the current prompt, the model-generated text, and, for content integration, an interactive knowledge-graph view. In the platform paper, this workflow is presented as a mechanism for reducing planning burden while retaining teacher oversight.

6. Knowledge integration and collaborative review

The platform’s knowledge integration pipeline is centered on knowledge-graph construction and retrieval-augmented prompting. The KG pipeline proceeds through schema design for the Content, Normative, and Evaluation KGs; triplet extraction via dependency parsing plus LLM semantic analysis; and triplet fusion across sources to align synonymous concepts from different disciplines, exemplified by “energy transfer” in physics and “heat flow” in chemistry. Fusion relies on graph-matching heuristics, specifically common predicates and shared root nodes, together with vector-similarity thresholds in the Neo4j embedding index (Wang et al., 3 Oct 2025).

The paper formalizes triplet extraction with the operation

$E_{\rm tab}$4

where $E_{\rm tab}$5 denotes annotated lesson-plan text. The accompanying pseudo-code splits the text into paragraphs, formats prompts of the form “Extract $E_{\rm tab}$6 from:”, calls the LLM on each paragraph, parses triplets from the response, and aggregates them into $E_{\rm tab}$7. Enhanced retrieval prompt engineering then retrieves the top-$E_{\rm tab}$8 most relevant case fragments using document vectors and appends KG-derived context to construct composite prompts.

Stepwise prompt generation is described using prompt-template fragments $E_{\rm tab}$9, user interdisciplinary parameters $d$0, intermediate drafts $d$1 through $d$2, a template-generation function $d$3, a prompt-optimization function $d$4, and an LLM generation function:

$d$5

$d$6

$d$7

$d$8

The review process is explicitly human–computer collaborative. After initial generation, teachers may edit prompts, insert domain-specific terminology, and resubmit revised prompts for re-drafting. Each iteration is logged, and the system stores prompt/version pairs for rollback and comparison. Several pedagogical rules are enforced within this loop. Alignment checks against the Normative KG require an activity’s stated learning outcome to appear as a node in the relevant curriculum-standard subgraph; otherwise, the platform issues a warning. Activity difficulty is cross-referenced against the Content KG’s “cognitive level” attribute, including Bloom’s taxonomy tags, and activities flagged as too advanced trigger a suggestion to scaffold via lower-level sub-nodes. A consistency rule requires every activity to link back to at least one learning objective node, thereby operationalizing backwards design.

7. Evaluation, implications, and limitations

The platform paper reports a two-phase within-subjects experiment with 43 teachers. The procedure consisted of a one-hour platform orientation and hands-on tutorial followed by small-group interdisciplinary lesson-plan creation tasks, with groups of 4–5 and multiple topics. Data collection included a pre-study Likert survey of prior experience, a post-study UTAUT-derived acceptance and behavioral intention questionnaire on a 5-point Likert scale, reflective journals, and system logs containing timestamps and prompt iterations (Wang et al., 3 Oct 2025).

Design efficiency was measured by total time to complete a standardized lesson-plan template. Lesson-plan quality was scored by the LLM-based evaluation module across 11 dimensions $d$9, such as rationality, interdisciplinarity, and challenge, with per-dimension scores defined as

$h_q$0

and overall quality defined as $h_q$1. Cognitive load and barriers were inferred from post-task reflective-journal coding rather than from a formal psychometric scale.

The paper reports a 75% efficiency increase and a 41% quality improvement. It states that the efficiency gain was computed from the difference between $h_q$2 and $h_q$3 relative to $h_q$4, and that the quality gain was computed from the difference between $h_q$5 and $h_q$6 relative to $h_q$7. It also reports UTAUT factors as follows:

Factor	Reported value
Performance Expectancy	$h_q$8
Effort Expectancy	$h_q$9
Facilitating Conditions	$h_t$0
Social Influence	$h_t$1
Behavioral Intention	$h_t$2

The study does not report inferential statistics such as $h_t$3-tests or confidence intervals for efficiency or quality scores. Reflective journals from 21 participants were coded into 98 advantage statements and 114 suggestions. The paper records teacher comments indicating reduced “confusion” and “mental fatigue” after the guided workflow and KG visualization became available, which it interprets as lower subjective cognitive effort.

The practical implications are framed around curriculum reform: teachers without strong interdisciplinary backgrounds can scaffold lesson plans more rapidly, the system can function as both an efficiency tool and a cognitive scaffold, and the platform is positioned relative to China’s new 10% interdisciplinary requirement. The stated limitations are equally explicit: scope is limited to compulsory education from K–9 and a small set of regional textbook versions; sample size and the absence of a control-group design limit generalizability and preclude formal statistical inference. Future directions include extending KG coverage vertically and horizontally, developing an “AI Teaching Assistant” that proactively flags alignment issues, incorporating multimodal generation such as diagrams, animations, and video snippets, building a teacher community for sharing and iteration, and expanding from “teaching design” to “learning design” with personalized student-side pathways.