Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
For the Kannada novel Heli Hogu Kaarana, Amruth Kumar M reports that dense-only retrieval returned the wrong chapter when asked, “What does Himavant say on page 120?” His revised system combined dense search with BM25 using Reciprocal Rank Fusion (RRF), and routed explicit page-number questions directly to page metadata. That is a useful design lesson from one book-sized project—not proof that hybrid search fixes Kannada retrieval in general.
Why dense-only retrieval missed the page-specific answer
Kumar built a retrieval-augmented generation (RAG) system around a scanned, 346-page physical copy of Ravi Belagere’s Kannada novel Heli Hogu Kaarana. In his account, a question about what Himavant says on page 120 led the initial dense-only system to a passage from the wrong chapter. He describes multilingual embeddings as inconsistent with inflected Kannada names in this implementation; that observation should not be generalized to every multilingual embedding model.
The problem reflects two different kinds of matching. Dense retrieval can find passages related in meaning, even when a question paraphrases the text. But a page reference or a less common name can depend on exact surface forms and reliable page provenance. A semantically related passage from elsewhere in a book is not a correct answer to a question about a particular page.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Kumar’s summary of that example was: “The difference was entirely in retrieval architecture.” That is his conclusion about this example, not a universal finding about RAG systems.
#1 Best Overall
How dense search, BM25 and RRF complement one another
| Method | What it contributes in this project | Where it can fall short |
|---|---|---|
| Dense retrieval | Finds semantically related passages and is reported as stronger for paraphrase and thematic questions. | May miss exact names or rare terms when the text form and query do not align well. |
| BM25 | Matches lexical surface forms and is reported as stronger for exact names and rare terms. | Does not inherently match different inflected forms. Kumar notes that ಹಿಮವಂತನ and ಹಿಮವಂತ are different tokens unless stemming or another morphological strategy is added. |
| RRF fusion | Combines dense and BM25 rankings by rank position, using k=60 in Kumar’s implementation. | Combines candidate rankings; it does not create a missing lexical or morphological match by itself. |
| Page metadata lookup | Routes explicit page-number questions to the requested page rather than asking semantic retrieval to infer it. | Depends on correct OCR and page metadata to identify and cite the right text. |
RRF is useful here because dense and BM25 scores are on different scales. Rather than adding those raw scores directly, it combines the rank positions each method assigns. A later reranking step can then reorder a narrower candidate set.
How the reported pipeline was built
The retrieval design depends on the document-processing stages before search. Kumar reports rendering PDF pages at 300 DPI, denoising and thresholding images with OpenCV, running Surya OCR for Kannada, applying Indic NLP normalization for Unicode and ligature cleanup, and semantically chunking the text with page metadata attached.
That page metadata serves two purposes: it supports direct routing for explicit page questions, and it helps preserve page provenance for citations. It is only as dependable as the OCR and metadata that produced it.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →For semantic or thematic questions
The system searches dense vectors in ChromaDB alongside BM25, merges the rankings with RRF at k=60, and reranks a reduced candidate set with mmarco-mMiniLMv2-L12-H384-v1. A conservative confidence threshold can refuse an answer that is not sufficiently grounded.
For explicit page-number questions
A regex route detects an explicit page reference and sends the request to a page metadata lookup instead of semantic search. This avoids asking a similarity model to determine which passage belongs to a page the user already named. Generation and optional speech follow later in the reported stack.
What the reported evaluation establishes—and what it does not
Kumar reports evaluating the project with a 50-query golden set used as a regression suite. The author gives these scores:
| Reported measure | Project result |
|---|---|
| Faithfulness | 0.92 |
| Answer relevancy | 0.88 |
| Context recall | 0.89 |
| Context precision | 0.85 |
These are figures reported by Kumar for this project, not independently reproduced results. The author explicitly says the 50-query set is not statistically powered. The reported exact-page route precision is 100% conditional on correct OCR and page metadata. Fallback model tiers were spot-checked for availability and quality trade-offs, not benchmarked. The figures therefore describe one system and corpus; they do not predict performance on another book, language, or deployment.
Latency and operational trade-offs
For the reported deployment, Kumar gives warm-path end-to-end latency of P50 2.8 seconds and P95 5.0 seconds. Explicit page routing is reported at approximately 5–12 milliseconds. The author also reports that serverless cold starts add approximately 1–2 seconds.
The project uses local ChromaDB, which the author identifies as a source of cold-start issues in serverless deployments. The broader stack includes local CPU retrieval and model-generation and text-to-speech calls. These figures are author-reported project measurements, not general service-level expectations; reranking also adds work after candidate retrieval.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.When this design is a reasonable fit
- Use dense retrieval when users mainly ask paraphrased or thematic questions and relevant passages may not share the query’s wording.
- Add BM25 when exact names, rare terms, or surface-form matches matter alongside semantic similarity.
- Use RRF to combine the two ranked lists without treating their raw scores as directly comparable.
- Route exact page questions separately when page identity is part of the request and trustworthy page metadata is available.
- Add language-aware normalization or morphology handling if inflected forms need to match; lexical BM25 alone does not bridge them in the reported implementation.
- Keep an abstention path for weakly grounded answers rather than forcing a response from low-confidence retrieval.
Limits readers should keep in view
This is a field report about one novel and one project evaluation, not an independent benchmark or a dataset audit. The author notes that OCR quality constrains provenance, feedback storage is ephemeral, and the system covers one book. Component-level retrieval scores are not reproduced in the article’s reported results, so the comparative strengths described here should be understood as the author’s account of this implementation.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
