Organizations: Department of Heritage Science & Technology, Indian Institute of Technology Hyderabad, Sangareddy, Telangana, India. · HTSR Institute, Bengaluru, Karnataka, India. · Bheri Technology Foundation, ITIC, Indian Institute of Technology Hyderabad, Sangareddy, Telangana, India. · Department of Biomedical Engineering, Indian Institute of Technology Hyderabad, Sangareddy, Telangana, India. · IBM (India), Bengaluru, Karnataka, India. · Central Sanskrit University, New Delhi, India.
Relevant evidence in Indian Knowledge Systems (IKS) collections may be distributed across passages in different parts of a work, including verses, translations, and commentaries. Locating it requires matching natural-language questions and partially remembered verse fragments to source passages. We present Dhara, a retrieval engine combining hybrid search, chapter- and book-level Entity Notes (EN), Fuzzy Verse Retrieval (FVR), and neural reranking. Five comparisons on the 325 questions in Dhara Bench examine retrieval representations against BM25, EN's contribution, reranker quality and computation time, FVR, and Gemini. BGE-M3 single-vector retrieval exceeds BM25 in nDCG@10, Recall@20, and MRR@20 across all categories. Its scores are broadly comparable to ColBERTv2, with mean retrieval times of 0.108 s and 0.962 s under the same hardware configuration and a smaller memory footprint for chunk storage on disk. Adding EN to a direct branch with a matched candidate count improves nDCG@10 in all twelve categories with Qwen and ten with BGE; the combined ranking is capped at the same candidate count. Qwen improves nDCG@10, Precision@5, and MRR@20 over BGE in every category, at approximately 6.72 times the overall estimated serial computation time. FVR increases Verse-category nDCG@10 by approximately 20 percentage points with either reranker. Gemini is evaluated after source mapping and manual review with relevance annotations visible. The Qwen pipeline with EN and FVR has higher Recall@20 than Gemini in ten categories and stronger complete-evidence retrieval in Multi-hop questions, while Gemini retains early-relevance advantages in several categories. The results inform retrieval choices according to evidence requirements and resource budgets.
Figures & tables
Approach
Derived retrieval representation
Organization / scope
Use of retrieved material
Dense X Retrieval Chen et al. (2024b)
Self-contained propositions
Fine-grained retrieval units
Retrieved propositions supply information for downstream tasks
RAPTOR Sarthi et al. (2024)
Recursive abstractive summaries
Tree spanning multiple abstraction levels
Chunks and summaries provide context for answer generation
GraphRAG Edge et al. (2024)
Community summaries
Communities in an entity graph
Community summaries support partial responses combined into an answer
EXPLAIN Liang et al. (2025)
Entity summaries
Summaries of linked entities
Summaries and document chunks provide retrieved answer context
EnSI-RAG Meng et al. (2026)
Structured entity records
Query-independent entity index
Records identify original passages used for answer synthesis
Dhārā
Entity Notes (EN)
CEN + BEN
Notes identify source chunks for reranking and evaluation against Dhārā Bench annotations
Table 1. Architectural comparison of generated retrieval representations and Entity Notes.
Figure 1. Component pipeline. Both branches and the combined ranking retain Sq candidates before the final cutoff and optional FVR.
Figure 2. Entity Note construction and retrieval. CEN records are consolidated into BEN records; retrieved notes identify linked source chunks for reranking.
Record type
Count
Source chunks
60,121
Chapter-level Entity Notes (CEN)
241,974
Book-level Entity Notes (BEN)
20,183
Parent verses
165,101
Generated verselets
715,158
Verse source records
15
Table 2. Retrieval-store inventory.
Figure 3. Verse-search lengths in Roman Colloquial (23,146 searches; one-character bins). Left: lengths through 60 characters. Right: the full 2-200-character range, with logarithmic frequency.
Figure 4. Experiment 1 compares direct retrieval with BM25 and informs the representation choice for Experiment 2. Four component comparisons examine EN, rerankers, FVR, and Gemini.
Code
Question category
n
T01
Fact Retrieval
51
T02
List / Set Retrieval
40
T03
Counting / Aggregation
22
T04
Chronology & Sequence understanding
25
T05
Lexical Understanding
31
T06
Composite
19
Table 3. Question categories in the 325-question Dhārā Bench evaluation set.
Configuration
Candidate selection and ranking
FVR
Direct matched+B/Q
Sq directly retrieved chunks, reranked
Off
EN+Direct+B/Q
Top Sq chunks after CombMAX of the EN and direct rankings
Off
EN+Direct+B/Q+FVR
Same merged ranking, with FVR matches prepended
On
Gemini
Generated references mapped to chunks and manually reviewed
-
Table 4. Component configurations. B/Q denotes BGE/Qwen reranking; Sq is the EN-derived candidate count.
Figure 5. Gemini as an LLM-based evidence retriever: source mapping with Gemma 4 31B IT, manual review with relevance annotations visible, and scoring at preserved result ranks.
Figure 6. Direct retrieval by question category: BGE-M3 and all-MiniLM-L6-v2 SV, ColBERTv2 MV, and BM25.
Figure 7. Experiment 1 overall scores across nine retrieval cutoffs; each bar averages all 325 questions.
Category
Shared
BGE-M3 only
ColBERTv2 only
T01
115
68
89
T02
56
40
67
T03
29
21
31
T04
34
28
25
T05
36
31
45
T06
23
12
13
Table 5. Relevant source chunks shared by BGE-M3 and ColBERTv2 or unique to each in the top-20 rankings, by category. Counts are question-chunk pairs.
Figure 8. Experiment 1 mean direct retrieval time across 325 questions, using the same hardware configuration and timing scope for all three models (Section 5.6 ).
Model
Representation
Memory footprint (disk)
all-MiniLM-L6-v2
Single vector
404.8 MiB
BGE-M3
Single vector
568.3 MiB
ColBERTv2
Multivector
7.30 GiB
Table 6. Memory footprints for chunk storage on disk: 60,121 source chunks per collection, including text, embeddings, indexes, and logs.
Figure 9. EN+SV minus SV matched, with FVR off. EN+SV and SV matched each retain Sq candidates. Score pairs show SV matched to EN+SV; bars show changes in percentage points (pp). Bottom panels report complete evidence.
Figure 10. BGE and Qwen retrieval scores versus estimated serial pipeline time for EN+SV, with FVR off. Lines connect the same question category; input pools and retained candidate counts are matched.
Component
Unit
BGE
Qwen
Embedding
ms
29.411
29.411
EN retrieval
ms
25.750
25.750
EN mapping
ms
19.323
19.323
SV retrieval
ms
241.616
241.616
EN reranking
s
2.806
19.700
SV reranking
s
3.336
23.395
Table 7. Mean component times across all 325 questions for EN+SV with FVR off. Each question has equal weight; units are shown per component.
Figure 11. FVR enabled and disabled in T08, holding the merged EN+SV ranking fixed before FVR.
Figure 12. EN+SV with FVR versus Gemini, for BGE and Qwen. Score pairs show Gemini to EN+SV+FVR; bars show changes in percentage points (pp). Complete-evidence metrics apply to T06 and T07.
Department of Heritage Science & Technology, Indian Institute of Technology Hyderabad, Sangareddy, Telangana, India. · HTSR Institute, Bengaluru, Karnataka, India. · Department of Biomedical Engineering, Indian Institute of Technology Hyderabad, Sangareddy, Telangana, India. +2