ChronoRAG Base: Scalable Temporal Retrieval

• ChronoRAG Base (CRB) is a hierarchical, tree-structured retrieval substrate that efficiently organizes heterogeneous time-series data across multiple domains.
• It employs a unified Apache Arrow data protocol and recursive k-means clustering to ensure scalable storage and domain-aware filtering with incremental update support.
• CRB integrates with the Multi-grained Series Interaction Learner and Model Cooperation Coherer in QuiZSF, enhancing forecast accuracy through external sequence knowledge injection.

ChronoRAG Base (CRB) is the retrieval substrate of QuiZSF, a zero-shot time-series forecasting framework that combines retrieval-augmented generation with time-series pre-trained models (TSPMs). In the formulation of QuiZSF, CRB is a hierarchical, tree-structured temporal database designed for scalable storage and domain-aware retrieval of time series, so that external sequence-level knowledge can be injected into forecasting models on demand. It is built from 27 datasets spanning seven domains—Energy, Health, IoT, Nature, Web, Transport, and Environment—and is released in three scales: Small with 34M points, Medium with 48M points, and Large with 143M points; the smaller versions are strict subsets of the larger ones (Ma et al., 9 Aug 2025).

1. System role and design objectives

Within QuiZSF, CRB serves a narrowly defined but structurally central function: it retrieves relevant historical sequences for a target series under zero-shot forecasting conditions. The architecture addresses two stated objectives. The first is efficient and domain-sensitive storage and retrieval over a large heterogeneous corpus of time series. The second is the injection of external knowledge into downstream TSPMs, where retrieved sequences are processed by the Multi-grained Series Interaction Learner (MSIL) and then integrated by the Model Cooperation Coherer (MCC) into both Non-LLM-based and LLM-based forecasters (Ma et al., 9 Aug 2025).

This placement is important for interpreting CRB correctly. It is not itself a forecasting model; rather, it is the retrieval base that supplies auxiliary context. The retrieved sequences do not bypass the learned forecasting stack. Instead, they are transformed into relational representations and then fused with the target sequence through the subsequent modules. A plausible implication is that CRB is best understood as a retrieval-and-organization layer for temporal knowledge, analogous in role to a corpus index in text RAG systems, but adapted to domain-partitioned sequence data.

2. Data protocol, normalization, and storage model

CRB adopts a unified data protocol intended to support heterogeneous time-series sources while preserving retrieval efficiency. All sequences are stored in Apache Arrow (ARROW) format. For a multivariate time series $\mathbf{X} \in \mathbb{R}^{N \times D}$ with $N$ time steps and $D$ channels, the $j$-th channel is denoted $\mathbf{x}_j = [x_{1j}, x_{2j}, \cdots, x_{Nj}]$. With window size $w$ and step size $s$ where $w \leq N$, sliding-window segmentation produces channelwise segments

$$\mathbf{X}_{kj} = [x_{(k - 1)s + 1,j}, \cdots, x_{(k - 1)s+w,j}],$$

subject to $(k - 1)s + w \leq N$ (Ma et al., 9 Aug 2025).

The database uses channel-independent processing,

$N$0

so that each dimension is processed independently. The paper motivates this as a simplification for database construction and fusion, and as a means to improve scalability and cross-domain adaptability. Missing values are completed by linear interpolation. Metadata are standardized through a key-value protocol with fields Domain/Category, Item_id, Start, End, Freq, and Target. The paper gives an example record with values such as Nature, us_births_dataset_0_0, 20000101, 20010527, Daily, and a target sequence beginning with $N$1.

Component	Content
Storage format	Apache Arrow (ARROW)
Segmentation	Sliding-window over each channel
Metadata fields	Domain/Category, Item_id, Start, End, Freq, Target

The sources used to build CRB are Monash, UCR, TSER, TDBrain, and UTSD. The use of a common schema across these sources is significant because cross-domain retrieval depends on consistent representation at storage time. This suggests that the “base” in ChronoRAG Base is not merely a collection of series, but a standardized temporal corpus designed to support retrieval under domain heterogeneity.

3. Hierarchical series tree and prototype organization

CRB is organized as a hierarchical series tree. At the top level, the full corpus is partitioned by domain:

$N$2

This domain partitioning provides the first stage of domain-aware filtering. Within each domain $N$3, recursive $N$4-means clustering is applied until each cluster contains at most $N$5 time series. The number of clusters for a domain is set as

$N$6

The clustering objective is

$N$7

where $N$8 is the centroid of the assigned cluster (Ma et al., 9 Aug 2025).

Each cluster stores a prototype chosen as the sequence closest to the centroid. These prototypes serve as representatives for approximate search, so the system can first compare a query to cluster representatives before scanning candidate clusters in full. The indexed elements specified in the paper are raw sequences and cluster prototypes, including centroids and nearest sequences. The paper does not specify indexing of explicit statistical features such as trend or seasonality descriptors, ACF/PACF, frequency-domain representations such as Fourier or wavelets, or learned embeddings beyond prototypes.

This point matters because one possible misconception is that CRB is a feature-engineered retrieval system. The paper does not support that interpretation. Retrieval is described as sequence-based and prototype-mediated, not as ANN over learned embeddings or as search over handcrafted signal descriptors. Conceptually, the hierarchy combines semantic partitioning by domain with geometric approximation by recursive clustering.

4. Retrieval mechanism and domain-aware matching

CRB retrieval is described through a domain-aware hierarchical procedure referred to in the paper as HHTR. Given a target time series $N$9, the objective is to retrieve Top-$D$0 relevant neighbors from the prototype database. Similarity is computed by a hybrid score that combines cosine similarity with Euclidean distance:

$D$1

where $D$2 is cosine similarity and $D$3 is Euclidean distance. The paper does not include DTW or alternative retrieval metrics (Ma et al., 9 Aug 2025).

If the domain of the target is known and present in CRB, retrieval combines local-domain and global matches:

$D$4

with $D$5. If the domain is unknown or absent, retrieval defaults to global matching only. Operationally, the query is first matched against cluster prototypes in a top-down traversal through domain groups and clusters, after which only a small number of candidate clusters are examined in detail.

The paper contrasts this empirically favorable behavior with the limitations of asymptotic claims. Flat linear retrieval has cost $D$6, and while tree structures are often associated with average-case $D$7 behavior, the paper explicitly cautions that such guarantees do not necessarily hold in high-dimensional settings because of the curse of dimensionality. No explicit numerical complexity or pseudocode is provided. A plausible interpretation is that CRB should be viewed as an empirically efficient retrieval structure rather than as a structure with a fully characterized theoretical complexity profile.

5. Coupling with MSIL and MCC

After retrieval, CRB hands a target sequence and its retrieved neighbors to MSIL. The inputs are a target sequence $D$8 and a retrieved set $D$9. The retrieved set is partitioned by domain as

$j$0

Target and retrieved sequences are normalized using the base-model scaler:

$j$1

MSIL then computes two forms of pattern representation. The interaction pattern and average pattern are defined as

$j$2

These are fused with the target by cross-attention:

$j$3

$j$4

MCC provides the model-side integration. For Non-LLM-based TSPMs, the Numerical Coherer applies residual fusion,

$j$5

For LLM-based TSPMs, the Language Coherer converts MSIL outputs $j$6 into structured textual summaries combined with an instruction-style prompt. Training uses an MSE objective regularized by MMD:

$j$7

In system terms, CRB therefore does not inject raw retrieved series directly into the predictor. It injects them through learned relational abstractions, which is a central design feature of QuiZSF rather than an auxiliary detail (Ma et al., 9 Aug 2025).

6. Empirical behavior, scaling effects, and usage characteristics

The paper reports ablations that isolate the effect of retrieval. Removing RAG from QuiZSF$j$8 causes a notable performance drop relative to the full system: on ETTm1 $j$9 ETTh2, Top-8 averaged prediction-length MSE changes from 0.382 to 0.394; on ETTm1 $\mathbf{x}_j = [x_{1j}, x_{2j}, \cdots, x_{Nj}]$0 ETTm2, it changes from 0.281 to 0.296. The paper further states that removing MSIL or the Coherer also hurts performance, but that the largest drop is observed for the w/o-RAG condition, which supports the importance of CRB within the full architecture (Ma et al., 9 Aug 2025).

Several hyperparameters govern CRB usage. CRB_Small performs worst, while larger CRB versions provide more external knowledge and better zero-shot forecasting performance; the paper describes this as consistent with scaling law observations. For retrieved neighbor count, TopK $\mathbf{x}_j = [x_{1j}, x_{2j}, \cdots, x_{Nj}]$1 yields the best results; retrieving more sequences may introduce noise, whereas fewer may provide less information. For the local-prototype proportion, a balanced setting of $\mathbf{x}_j = [x_{1j}, x_{2j}, \cdots, x_{Nj}]$2 gives a good trade-off, while overly large same-domain emphasis reduces diversity and may lead to polarized outcomes.

The efficiency claims are empirical rather than fully quantified in the text. The paper reports model-size and CPU inference-time comparisons per batch and concludes that QuiZSF$\mathbf{x}_j = [x_{1j}, x_{2j}, \cdots, x_{Nj}]$3 introduces only minor overhead from retrieval and interaction modules while remaining lightweight. Exact numerical memory footprint and latency values are not provided in the text. At the full system level, QuiZSF with Non-LLM-based and LLM-based TSPMs ranks Top1 in 75% and 87.5% of prediction settings, respectively, while maintaining high efficiency in memory and inference time.

A concrete usage example is given for ETTh2. HHTR retrieves Top-8 similar sequences from CRB; after normalization and MSIL processing, the resulting retrieved patterns help smooth noisy predictions and reduce hallucinations. The paper attributes smoother forecasts to RAG via CRB because the retrieved patterns reveal average trends, whereas outputs without RAG fluctuate more and contain more inaccurate details.

7. Incremental updates, limitations, and open questions

CRB supports incremental updates. A new sequence is assigned to the nearest prototype and inserted into the matched cluster. If the cluster size exceeds $\mathbf{x}_j = [x_{1j}, x_{2j}, \cdots, x_{Nj}]$4, local re-clustering is triggered by re-optimizing the affected subtree rather than rebuilding the entire database. This local overflow handling is part of the stated scalability design.

The paper also places clear boundaries on what CRB presently does not provide. It does not provide exact algorithms or pseudocode for the hierarchical retrieval procedure. It does not provide explicit complexity figures for CRB operations. It does not compare retrieval quantitatively against FAISS, HNSW, or other ANN libraries. It does not specify indexing of ACF/PACF, Fourier or wavelet features, or learned embeddings beyond raw sequences and prototypes (Ma et al., 9 Aug 2025).

Observed limitations are partly CRB-related and partly framework-level. QuiZSF$\mathbf{x}_j = [x_{1j}, x_{2j}, \cdots, x_{Nj}]$5 shows limitations on short-term, minute-level forecasting such as ETTm1, where aligning retrieved knowledge with fine-grained fluctuations is difficult. This suggests a retrieval-granularity mismatch when external sequences are too coarse relative to the target dynamics. More generally, the framework depends on the quality and relevance of pre-trained models, incurs complexity from large-scale retrieval, and may benefit from sparse connections in the retrieval database and stronger multi-granularity sequential learning.

Taken together, these constraints position CRB as a pragmatic retrieval infrastructure rather than a fully closed theoretical indexing framework. Its contribution is the combination of domain partitioning, recursive prototype clustering, and retrieval-mediated interaction with TSPMs. In QuiZSF, that combination turns a static time-series corpus into an external temporal memory for zero-shot forecasting.