Adaptive hybrid AI-based retrieval-augmented generation for construction and seismic regulatory question answering

3.1. System overview

The overall architecture consists of the following stages:

1) Document Processing (PDF extraction and cleaning).

2) Text Chunking and Representation.

3) Vector Indexing (FAISS).

4) Hybrid Retrieval Engine.

5) Adaptive Weighting Mechanism.

6) Response Generation (LLM).

Let the document corpus be defined as:

Each document is split into smaller chunks:

Similarly, a query $q$ is mapped into the same vector space:

3.2. Retrieval channels: semantic, keyword-based and rule-based scoring

Semantic similarity is computed using cosine similarity, defined in Eq. (5):

After L2-normalization, cosine similarity reduces to:

For the IndexFlatIP configuration, retrieval complexity is $O\left(n\right)$.

To improve lexical matching, we adopt the BM25 scoring function [29], given in Eq. (7):

where $f\left(t,c_i\right)$ is the term frequency, $\left|c_i\right|$ is chunk length, and $avgdl$ is the average document length.

For structured engineering domains such as construction and seismic regulations (QMQ and ShNQ), we define a weighted rule-based scoring function, Eq. (8):

where pattern weights are defined by Eq. (9) as:

The matching function is given by Eq. (10) as:

where $\delta \in \left(\mathrm{0,1}\right)$ represents partial matching weight.

The pattern set P comprises regular expressions reflecting the clause-numbering conventions of Uzbek regulatory documents: full document codes (e.g., QMQ 2.01.03-96), section and clause references, table and appendix identifiers, and numeric parameter expressions such as seismic intensity values. Structured queries containing such identifiers are precisely those for which the rule-based channel contributes most strongly; its quantitative impact on retrieval accuracy is isolated in the ablation study (Section 5.3).

3.3. Adaptive fusion and query-adaptive weighting

We define a non-linear scoring function to combine multiple retrieval signals, Eq. (11):

where $\sigma \left(\cdot \right)$ is the sigmoid activation function.

The relative influence of the three retrieval channels is assigned per query as:

where $\phi \left(q\right)$ is a vector of binary indicators obtained by thresholding simple query statistics – the presence of a document-code or clause pattern, an indicator of high digit density, and an indicator of long queries – with the indicator thresholds recorded in the released configuration, and $W$ is a fixed, expert-defined mapping matrix that raises the rule-based weight for structured queries and the semantic weight for natural-language queries. In the current implementation $W$ is not trained; learning it from annotated query logs is identified as future work (Section 7). The interaction coefficient $w_4\ge 0$ in Eq. (11) is a fixed, non-negative configuration constant chosen together with the mapping $W$; it is not produced by the softmax of Eq. (12) and is therefore treated as a separate bounded parameter in the analysis of Section 3.4.

3.4. Analytical properties of the fusion model

The fused score of Eq. (11) possesses several properties that make its behaviour predictable and analysable. Because the outer function is the logistic sigmoid $\sigma \left(z\right)=\frac{1}{1+e^{-z}}$, the fused score is bounded, $Score\in \left(\mathrm{0,1}\right)$, for any values of the channel scores, so that no single channel can produce an unbounded contribution to the ranking. Differentiating Eq. (11) with respect to each channel score yields the partial derivatives given in Eq. (13):

The bilinear interaction term $w_4·S_{semantic}·S_{rule}$ acts as a conjunctive amplifier: its contribution is significant only when the semantic and the rule-based channels return high scores simultaneously, and it vanishes whenever either factor is close to zero. For structured queries this rewards chunks that are both topically close to the query and contain the exact clause identifier – precisely the behaviour required in regulatory retrieval – whereas for natural-language queries, where $S_{rule}\approx 0$, the model degrades gracefully to a weighted combination of semantic and lexical evidence.

The query-adaptive weights of Eq. (12) are produced by the softmax operator and therefore satisfy $w_i>0$ and $\sum w_i=1$ for every query, which fixes the relative budget of the three linear terms. The channel scores themselves, however, live on different scales – a signed cosine, a non-negative BM25 score bounded by $B_q$ and a non-negative rule score bounded by $W_P\left(q\right)$ – so the pre-activation is not a convex combination of commensurate quantities; what the analysis below relies on is the weaker but sufficient property that it is bounded for every query, as summarised in Eq. (14).

The keyword channel inherits the saturation behaviour of BM25: for a fixed document length, the contribution of a term grows with its frequency but is bounded by the limit $(k_1+1)·IDF\left(t\right)$ as the frequency tends to infinity, so that $k_1$ controls how quickly repeated occurrences of a term stop adding evidence, while $b\in \left[\mathrm{0,1}\right]$ interpolates between no document-length normalisation ($b=0$) and full normalisation ($b=1$). These are exactly the two parameters, together with the partial-match weight $\delta$ of Eq. (10), whose ranges are identified in Section 5.3.

Two further properties are worth stating. First, because the sigmoid is strictly increasing, the ranking of candidates induced by the fused score coincides exactly with the ranking induced by the pre-activation $z$ itself; the non-linearity therefore does not alter the retrieval order but maps the score into (0, 1), the scale on which the refusal mechanism of Section 4 operates. Because the bound of Eq. (14) is query-dependent (through $B_q$ and $W_P\left(q\right)$), large lexical scores can drive the sigmoid close to saturation; the numerical value of the refusal threshold is therefore a configuration parameter chosen on a development set rather than a universal constant. The roles remain cleanly separated: the linear-plus-interaction form of $z$ determines which chunks are returned, while $\sigma$ provides the common (0, 1) scale on which the answering decision is made.

Second, all three channel scores are bounded for every fixed query, as summarised in Eq. (14): with the $L_2$-normalised embeddings of Eq. (6) the semantic score is a cosine similarity in [–1, 1]; the BM25 score is bounded by the finite query-dependent constant $B_q$; and the rule score is bounded by the total weight $W_P\left(q\right)$ of the patterns matched in the query, so that the pre-activation $z$ is finite for every query. Small perturbations of the evidence therefore produce proportionally small changes of the score: since $\sigma \text{'}\left(z\right)\le \frac{1}{4}$, Eq. (13) implies that the fused score is Lipschitz-continuous in each channel, with the constants $\frac{w_2}{4}$ for the keyword channel, $\frac{w_3+w_4}{4}$ for the rule channel and $\frac{w_1+w_4·W_P\left(q\right)}{4}$ for the semantic channel:

The keyword bound is defined through absolute IDF values, so that it holds for any IDF variant – including formulations in which the IDF of very frequent terms may become negative – while the rule bound equals the total weight of the patterns matched in the query.

Finally, the computational cost of the fusion stage is negligible compared with candidate generation. Exact inner-product search over the $n$ candidate vectors presented to the search stage, each of dimension $d$, costs $O(n·d)$ per query (Section 4); BM25 scoring is linear in the total number of postings of the query terms; and rule matching is linear in the number of patterns and the query length. The fusion itself is $O\left(1\right)$ per candidate, so the overall per-query complexity is dominated by the index search, which is consistent with the latency figures reported in Section 6.

These observations can be summarised as six properties: (P1) the fused score is bounded in (0, 1); (P2) it is strictly monotone in the semantic and keyword channels, and in the rule channel under the explicit sign condition stated above; (P3) the interaction term implements a conjunction of semantic and rule evidence; (P4) the softmax normalisation fixes the budget of the linear terms ($w_1+w_2+w_3=1$), with the fixed interaction coefficient $w_4$ treated as a separate bounded parameter; (P5) the candidate ranking is invariant to the sigmoid and is decided by the pre-activation alone; and (P6) the score is Lipschitz-continuous with explicit per-channel constants. Together, P1-P6 characterise the fusion model as a bounded, stable and order-preserving scoring rule whose behaviour can be predicted analytically for any query type – a level of transparency that purely learned, end-to-end retrieval scorers do not generally provide.

Proposition 1 (properties of the fused score). For every query $q$ and candidate chunk $c$: (i) $Score\in \left(\mathrm{0,1}\right)$; (ii) Score is strictly increasing in $S_{semantic}$ and $S_{keyword}$; it is strictly increasing in $S_{rule}$ whenever $w_3+w_4·S_{semantic}>0$ — in particular whenever $S_{semantic}\ge 0$, for every $S_{semantic}\in [-\mathrm{1,1}]$ when $w_4<w_3$, and for all cases when $w_4=0$; (iii) the ranking of candidates induced by Score coincides with the ranking induced by the pre-activation $z$; (iv) the pre-activation is bounded for every query as stated in Eq. (14); and (v) Score is Lipschitz-continuous in each channel with the explicit constants stated above. Proof: (i) follows from $\sigma \in \left(\mathrm{0,1}\right)$; (ii) from $\sigma \text{'}\left(z\right)>0$, the derivatives of Eq. (13), $S_{rule}\ge 0$ and the stated sign condition for the rule channel; (iii) from the strict monotonicity of $\sigma$; (iv) from the channel bounds $S_{semantic}\in [-\mathrm{1,1}]$, $S_{keyword}\le B_q$ and $S_{rule}\le W_P\left(q\right)$; and (v) from $\sigma \mathrm{\text{'}}\left(z\right)\le \frac{1}{4}$ applied to the derivatives of Eq. (13). □

3.5. Retrieval pipeline and answer generation

The proposed multi-stage retrieval pipeline operates as follows:

1) Rule-based pre-filtering using $S_{rule}$.

2) BM25-based keyword filtering using $S_{keyword}$.

3) FAISS-based semantic retrieval using $S_{semantic}$.

This hierarchical filtering reduces the effective search space and improves practical retrieval efficiency. These numbered stages describe the order of computation: the earlier, inexpensive filters successively narrow the candidate pool, and the three channel scores entering the fusion of Eq. (11) are then evaluated for the candidates that survive filtering. Fig. 1 therefore depicts the logical scoring relationship among the channels, while the list above describes their sequential execution; the two views are complementary descriptions of the same pipeline rather than alternative algorithms. A stage acts as a filter only when it fires: if no rule pattern matches the query, the rule stage leaves the candidate set unchanged and simply contributes a zero rule score to Eq. (11), and the keyword stage behaves in the same way for query terms absent from the index, so the pipeline degrades gracefully to the remaining channels.

Top-$k$ retrieved chunks are used as context, Eq. (15):

The final response is generated using a large language model, Eq. (16):

This follows the RAG paradigm while enhancing retrieval robustness and efficiency [1, 2].

The augmentation step itself is performed deterministically. The prompt supplied to the generative model is assembled from three components: a fixed system instruction that restricts the model to the supplied context and requires clause identifiers to be cited; the top-$k$ retrieved chunks, ordered by decreasing fused score and prefixed with their document and clause identifiers; and the user query. If the concatenated context exceeds the token budget of the model, the lowest-ranked chunks are truncated first, so that the strongest evidence is always preserved. Augmentation therefore modifies neither the model weights nor its decoding procedure; it only conditions generation on retrieved, verifiable evidence, which is what distinguishes the augmented configuration analysed in Section 6 from a non-augmented (closed-book) use of the same model.

3.6. System architecture

This section describes the overall architecture of the proposed Adaptive Hybrid Retrieval-Augmented generation model (AHRM), highlighting the end-to-end data flow from PDF document ingestion to real-time API responses. The overall architecture is summarised in Fig. 1.

The indexing pipeline instantiates the formalism of Section 3 directly: the corpus and chunk structure follow Eqs. (1-2); chunks and queries are embedded into the shared vector space of Eqs. (3-4) using the multilingual Sentence-BERT model [28]; and L2-normalisation of the stored vectors reduces the cosine similarity of Eq. (5) to the inner product of Eq. (6), which is exactly the quantity computed by the FAISS index [30]. No additional formalism is therefore required at the implementation level.

Fig. 1Overall architecture of the proposed AHRM: offline indexing of QMQ/ShNQ documents and online multi-channel retrieval with adaptive fusion and LLM-based answer generation

Chunking preserves semantic integrity and table/formula context. In our implementation, a hybrid text splitter ensures that paragraphs, tables, and formulas remain within single chunks, preventing context fragmentation.

Normalized vectors are stored in a FAISS IndexFlatIP index for fast nearest neighbor search [30, 31]. The IndexFlatIP configuration performs exact (exhaustive) inner-product search, whose cost grows linearly with the number of stored vectors; at the current corpus scale (approximately 85,000 chunks) this remains fast in practice, while for substantially larger corpora approximate structures such as IVF or HNSW [30, 31] provide sub-linear scaling at a controlled loss of accuracy.

The proposed retrieval framework integrates semantic similarity, keyword-based matching, and rule-based filtering within a unified pipeline. Instead of combining these signals using a linear formulation, the model employs a non-linear scoring function (see Section 3.3) to capture interactions between semantic and structured features.

This design is particularly effective for structured engineering queries, where numerical identifiers (e.g., QMQ clause numbers), seismic parameters, and domain-specific terminology must be accurately matched. This approach improves retrieval robustness in safety-critical domains, where precise matching of regulatory constraints directly impacts decision reliability.

The top-$k$ retrieved chunks are served through a RESTful API, which is integrated with a ReactJS-based chatbot interface:

1. The user submits a query through the chatbot interface.

2. The API performs efficient retrieval and scoring of candidate chunks.

3. The top-$k$ ranked chunks are concatenated to form the context.

4. The constructed context is passed to the LLM (DeepSeek-R1 via Ollama) to generate the final response.

The DeepSeek-R1 model served via Ollama was selected for three practical reasons: it is an open-weight model that can be deployed entirely on local infrastructure, which is a strict requirement because regulatory documents cannot be transmitted to external cloud services; its reasoning-oriented architecture is well suited to multi-step regulatory queries; and it makes Uzbek- and Russian-language input feasible; a systematic assessment of its behaviour on domain-specific terminology is part of the planned extended evaluation (Section 7). A systematic comparison with alternative generative models is beyond the scope of this study and is identified as future work (Section 7).

To reduce the risk of hallucinated answers when retrieved chunks are only partially relevant, three safeguards are applied. First, the generation prompt explicitly instructs the model to answer strictly on the basis of the supplied context. Second, queries whose maximum fused retrieval score falls below a confidence threshold are answered with an explicit refusal rather than a speculative response. Third, generated answers cite the identifiers of the source clauses, allowing the user to verify each statement against the original document.

Since the chatbot operates as a networked, interactive client-server application, its deployment also takes into account security considerations discussed for interactive communication media algorithms in networking applications [38] and for secure data transmission in sensor-based IoT deployments [39].

Because the underlying corpus combines documents in Uzbek and Russian and user queries may mix both languages, a multilingual Sentence-BERT model is used for embedding, so that queries and chunks in different languages are mapped into a shared vector space; the keyword and rule-based channels operate on the surface forms present in the documents. A systematic evaluation of cross-lingual retrieval quality is left for future work (Section 7).

Fig. 2Real-time chatbot interface of the proposed AHRM system for construction and seismic regulatory question answering. The interface illustrates user query input, retrieval of relevant QMQ and ShNQ clauses, and generation of structured, context-aware responses

In our measurements, the index-search stage (candidate scoring and ranking across the FAISS, BM25 and rule channels) required approximately 0.6 ms per query on average over the filtered candidate pool produced by the pre-filtering stages of Section 3.5 (not an exhaustive scan of the full index) (top-$k$ results, Intel Core i7 CPU). This figure refers exclusively to index search: query embedding adds a further overhead of its own, and the end-to-end response time is dominated by LLM answer generation. From the user’s perspective the interaction remains real-time, demonstrating the practical deployment of the system for engineering regulatory assistance.

To illustrate the practical deployment of the proposed system, the chatbot interface is presented in Fig. 2.

As shown in Fig. 2, the proposed system supports real-time interaction by integrating fast retrieval with generative response synthesis. The interface highlights how relevant regulatory clauses are retrieved and utilized to generate structured, context-aware responses.

4.1. Dataset

We utilize a curated private corpus of Uzbek construction and seismic regulatory documents, including PDFs of the Constitution, laws, building codes (QMQ and ShNQ), regulations, and related normative documents. The dataset comprises:

– 120 PDFs (approx. 12,000 pages).

– 85,000 textual chunks after processing (tables, formulas, paragraphs).

– Diverse query set including structured regulatory references (article numbers, construction code references) and natural-language queries.

Note: This is a private dataset and is not publicly available. Access may be granted upon reasonable request for research purposes.

The dataset was preprocessed using pdfplumber for text extraction, table parsing, and formula retention. Chunks were generated using RecursiveCharacterTextSplitter, with a configured maximum chunk size of 800 and an overlap of 150, chosen to maintain semantic context; these values are expressed in the splitter’s configured length unit – characters in the reference implementation, tokens when a tokenizer-based length function is supplied – and the deployed setting is recorded in the released configuration.

In terms of composition, the corpus is dominated by construction norms and rules (QMQ) and urban-planning norms (ShNQ), complemented by the primary legislation that frames them; the documents combine running text with tables, formulas and numbered clause hierarchies, and are written partly in Uzbek and partly in Russian. With the chunking configuration of Section 4 (a configured maximum chunk size of 800 with an overlap of 150), the 120 documents of approximately 12,000 pages yield approximately 85,000 chunks – on average about seven chunks per page – each associated with its source document and clause identifiers, which is what enables the clause-citing generation described in Section 4. The evaluation query set mirrors the two usage patterns observed in engineering practice: structured queries built around explicit code and clause references, and natural-language queries formulated without identifiers, in both Uzbek and Russian. The recursive splitter yields variable-length chunks bounded above by the configured maximum, splitting preferentially at paragraph and sentence boundaries so that semantic units are kept intact. Table 1 summarises the corpus and query-set statistics.

4.2. Evaluation metrics and baselines

To assess retrieval and QA performance, we used standard metrics in IR and NLP [27]:

1) Precision@$k$ ($P@k$): fraction of relevant chunks retrieved among the top-$k$ candidates.

2) Mean Reciprocal Rank (MRR): evaluates ranking quality of first relevant chunk.

3) Retrieval Latency ($T_{retrieval}$): time in milliseconds per query.

4) Answer Accuracy: measured as the percentage of generated responses that correctly match expert-validated ground truth answers.

Precision@$k$ and MRR reflect the top-$k$ question-answering setting, in which only the $k$ highest-ranked chunks are passed to the generative model. Recall@$k$ (the fraction of all relevant chunks for a query that appear among the top-$k$ candidates) and F1@$k$ (the harmonic mean of Precision@$k$ and Recall@$k$) characterise, in addition, the coverage of all relevant clauses; their systematic measurement over the full query set is part of the planned extended evaluation (Section 7).

Answer accuracy is assessed by expert judgement against the gold clauses identified for each query. In the planned extended evaluation, every point metric will additionally be accompanied by bootstrap confidence intervals computed over the query set, so that differences between configurations can be assessed statistically rather than through point estimates alone.

Formally, for a query $q$:

We compared AHRM against three baseline retrieval models:

1) Dense Retrieval (Sentence-BERT) [28]: standard FAISS-based semantic search.

2) Keyword Matching [29]: traditional term-based retrieval.

3) Hybrid Retrieval with Static Weights: combines dense and keyword retrieval with fixed weights.

All configurations – the three baselines above and the full AHRM – are executed within the same pipeline implementation and differ only in the channels enabled, so that differences in ranking quality and latency are attributable to channel composition rather than to implementation artefacts.

4.3. Ablation and sensitivity analysis

To evaluate the contribution of each component in AHRM, we conducted an ablation study:

– Removing keyword-based scoring decreased $P@3$ to 0.80.

– Removing rule-based scoring decreased $P@3$ to 0.77.

– Using static weights decreased $MRR$ to 0.73.

These results indicate that both adaptive weighting and rule-aware retrieval contribute substantially to overall system performance.

The robustness of the fused scoring function of Eq. (11) depends on its hyper-parameters: the BM25 parameters $k_1$ and $b$, and the partial-match weight $\delta$ of the rule-based channel. A systematic grid evaluation of Precision@3 and MRR over $k_1\in \left[\mathrm{0.9,1.8}\right]$, $b\in \left[\mathrm{0.5,1.0}\right]$ and $\delta \in \left[\mathrm{0.3,0.7}\right]$ is identified as part of the planned extended evaluation (Section 7).