vltruong01/TechQA-Hybrid-RAG-System-Comparison

TechQA Hybrid RAG System Comparison

A reproducible evaluation package comparing five retrieval strategies for technical-support Retrieval-Augmented Generation (RAG) on the reduced nvidia/TechQA-RAG-Eval corpus.

This repository contains the executed experiment notebook and evaluation artifacts used to compare lexical, dense, hybrid, reranked, and query-adaptive retrieval under a controlled downstream generation setup.

Primary retrieval result: Adaptive Hybrid RRF achieved the strongest overall held-out document-ranking performance and was selected as the primary retrieval strategy.

Important downstream result: Static Hybrid RRF remained highly competitive and achieved the strongest automated generation metrics on the fixed 30-question generation subset.

These are not contradictory results: document ranking quality and frozen-generator answer quality are related, but they are not the same objective.

---

Quick overview

Item	Setting
Upstream dataset	nvidia/TechQA-RAG-Eval
Experimental corpus	496 unique Technotes
QA rows	910
TRAIN	600 questions
DEV	310 questions
Answerable DEV retrieval evaluation	160 questions
Generation evaluation	Fixed 30 answerable DEV questions
Dense retriever	BAAI/bge-base-en-v1.5
Neural reranker	BAAI/bge-reranker-base
Generator	Qwen/Qwen2.5-7B-Instruct
Generator precision	4-bit NF4
Chunk size / overlap	384 / 64 tokens
Evidence budget	5 chunks, max 2 per document
Max prompt length	4096 tokens
Max new tokens	256
Fine-tuning	None
Primary retrieval strategy	Adaptive Hybrid RRF

---

1. What this repository is

This is an experiment / evaluation repository, not a newly trained language model and not a replacement copy of the upstream TechQA dataset.

It is intended to make the system comparison transparent and reproducible by publishing:

• the complete executed notebook;
• per-system retrieval metrics;
• per-question retrieval outputs;
• all 150 generated answers from the fixed generation comparison;
• Adaptive-router training and diagnostics;
• abstention / evidence-sufficiency analysis;
• Adaptive-vs-baseline pairwise summaries;
• experiment configuration and checksums.

What this repository is not

• It is not a fine-tuned Qwen checkpoint.
• It is not a fine-tuned BGE model.
• It is not a new version of TechQA.
• It does not claim that Adaptive Hybrid RRF wins every downstream metric.
• It does not claim statistical significance.

A separate Hugging Face Model repository for the selected Adaptive Hybrid RRF system will be linked here after it is published.

---

2. Research question

How effective are lexical, dense, static-hybrid, neural-reranked, and query-adaptive retrieval strategies for retrieval-augmented technical-support question answering when the corpus, chunking, generator, prompt, and evidence budget are held constant?

The controlled variable is the retrieval strategy.

The BGE retriever, reranker, and Qwen generator remain frozen. No LoRA or QLoRA is used.

---

3. Systems compared

System	Role	Main idea
BM25	Lexical baseline	Exact-term / token matching
Dense BGE	Semantic baseline	Dense similarity using BAAI/bge-base-en-v1.5
Static Hybrid RRF	Fixed hybrid baseline	Equal-weight BM25 + Dense Reciprocal Rank Fusion
Static RRF + Neural Reranker	Reranking ablation	Cross-encoder reranking of Static-RRF candidates
Adaptive Hybrid RRF	Primary retrieval strategy	TRAIN-learned query router changes BM25/Dense fusion weights per query

The neural reranker is an evaluation-only ablation. It is not part of the Adaptive Hybrid RRF pipeline.

---

4. Experimental design

The experiment keeps the corpus, chunking, generator, prompt, decoding settings, and evidence budget fixed across all five systems.

Reduced TechQA-RAG-Eval
        |
        +-- TRAIN
        |    |
        |    +-- query-router training
        |    +-- adaptive fusion-strength selection
        |    `-- evidence-sufficiency / abstention calibration
        |
        `-- DEV
             |
             +-- full retrieval evaluation
             +-- fixed 30-question generation comparison
             `-- qualitative answer inspection

Data split used by the experiment

Split	Total	Answerable	Impossible
TRAIN	600	450	150
DEV	310	160	150
Total	910	610	300

Chunking and evidence budget

• text-preserving chunking;
• chunk size: 384 tokens;
• overlap: 64 tokens;
• generation evidence budget: 5 chunks;
• maximum 2 chunks per document;
• maximum prompt length: 4096 tokens;
• maximum generated length: 256 new tokens.

---

5. Adaptive Hybrid RRF

Adaptive Hybrid RRF keeps the same BM25 and Dense BGE retrievers as the fixed hybrid baseline, but replaces a single global fusion ratio with a query-dependent fusion decision.

                    Question
                       |
             +---------+---------+
             |                   |
             v                   v
        BM25 ranking        Dense BGE ranking
             |                   |
             +---------+---------+
                       |
                       v
             TRAIN-learned router
                       |
                       v
          query-specific fusion weights
                       |
                       v
            Reciprocal Rank Fusion
                       |
                       v
            Adaptive Hybrid RRF

Router training

The router is trained only on answerable TRAIN questions.

TRAIN router target distribution:

Router class	Samples
BM25-heavy	17
Balanced	371
Dense-heavy	62
Total	450

Out-of-fold router diagnostics:

Metric	Value
Accuracy	0.7311
Balanced accuracy	0.5182
Macro F1	0.4762

The selected specialist fusion boost is:

ADAPTIVE_HEAVY_WEIGHT = 3.0

It is chosen from TRAIN out-of-fold evaluation, not from DEV.

The router is intentionally lightweight: it adapts the fusion weights without fine-tuning BGE or Qwen.

---

6. Held-out retrieval results

Retrieval is evaluated on all 160 answerable DEV questions.

System	R@1	R@5	R@10	MRR@10	nDCG@10	Mean latency (ms)
Adaptive Hybrid RRF	0.8500	0.9438	0.9688	0.8891	0.9083	54.63
Static Hybrid RRF	0.8500	0.9313	0.9688	0.8884	0.9076	57.07
Dense BGE	0.7938	0.9625	0.9750	0.8617	0.8899	20.01
BM25	0.8000	0.9063	0.9250	0.8473	0.8664	33.20
Static RRF + Neural Reranker	0.7438	0.9313	0.9688	0.8213	0.8572	1592.13

Retrieval interpretation

• Adaptive Hybrid RRF has the strongest overall top-rank performance by MRR@10 and nDCG@10 and ties for the best R@1.
• Dense BGE remains strongest at deeper R@5/R@10 recall.
• Static Hybrid RRF is extremely competitive and only slightly behind Adaptive on the ranking metrics.
• Neural reranking does not improve the controlled baseline and adds substantial latency.

For this reason, Adaptive Hybrid RRF is selected as the primary retrieval strategy, not as a claim that it dominates every metric.

---

7. End-to-end generation comparison

Generation uses the same frozen Qwen/Qwen2.5-7B-Instruct model, prompt design, decoding configuration, and 5-chunk evidence budget for every retrieval system.

The comparison uses the same fixed 30 answerable DEV questions for all five systems.

System	Answer rate	False abstention	Gold retrieval	ROUGE-L	BGE answer cosine
Static Hybrid RRF	60.0%	40.0%	96.7%	0.2249	0.6914
Adaptive Hybrid RRF	60.0%	40.0%	100.0%	0.2081	0.6790
Dense BGE	53.3%	46.7%	96.7%	0.1854	0.6743
BM25	46.7%	53.3%	90.0%	0.1855	0.6540
Static RRF + Neural Reranker	56.7%	43.3%	96.7%	0.1575	0.6691

Generation interpretation

The generation experiment shows an important distinction:

Best overall held-out document ranking
        -> Adaptive Hybrid RRF

Best automated generation metrics on the fixed 30-question subset
        -> Static Hybrid RRF

Adaptive retrieved the gold document for 30/30 generation questions, but the frozen generator still abstained on some answerable questions or did not always use the retrieved evidence optimally.

Static and Adaptive produced identical generated answers on most questions. The small number of differing outputs is one reason the modest generation-score difference should not be overstated.

Full outputs are published in:

results/generation_eval_outputs_comparison.csv

This file includes:

• question;
• reference answer;
• generated answer;
• gold-document retrieval flag;
• evidence similarity;
• abstention status;
• citation checks;
• ROUGE-L;
• BGE answer similarity;
• latency.

---

8. Evidence-sufficiency / abstention analysis

Retrieval relevance and answerability are treated as related but separate problems.

A lightweight evidence-sufficiency classifier is calibrated using TRAIN out-of-fold predictions.

TRAIN calibration

Item	Value
Selected threshold	0.535
TRAIN OOF ROC-AUC	0.7806

Held-out DEV evaluation

The auxiliary classifier is evaluated on all 310 DEV questions:

Metric	Value
Answerable answer rate	0.6750
Impossible abstention rate	0.8133
False-answer rate on impossible questions	0.1867
Balanced accuracy	0.7442
Overall accuracy	0.7419
ROC-AUC	0.8095

This module is reported as an auxiliary diagnostic, not as the main research contribution.

---

9. Why retrieval and generation can disagree

A correct document being retrieved does not guarantee that the generator will produce the best answer.

Possible downstream failure points include:

correct document retrieved
        |
        v
relevant passage may be split across chunks
        |
        v
only a limited evidence budget reaches Qwen
        |
        v
Qwen may under-use evidence or abstain
        |
        v
generated answer can still be incomplete or wrong

This explains why Adaptive can achieve the strongest full-DEV ranking while Static can still obtain slightly stronger generation metrics on a smaller fixed subset.

---

10. Repository contents

TechQA-Hybrid-RAG-System-Comparison/
|
|-- README.md
|-- TechQA_Hybrid_RAG_Qwen25_7B_System_Comparison.ipynb
|-- requirements.txt
|-- manifest.json
|
`-- results/
    |-- retrieval_metrics_comparison.csv
    |-- retrieval_detail_comparison.csv
    |-- generation_eval_questions_comparison.csv
    |-- generation_eval_outputs_comparison.csv
    |-- generation_metrics_comparison.csv
    |
    |-- adaptive_weight_search_comparison.csv
    |-- router_*.csv
    |-- abstention_*.csv
    |-- adaptive_vs_*.csv
    |-- oracle_specialist_upper_bound_comparison.csv
    |-- manual_generation_review_template_comparison.csv
    `-- experiment_config_comparison.json

Key files

File	What it contains
TechQA_Hybrid_RAG_Qwen25_7B_System_Comparison.ipynb	Complete executed experiment and reproduction notebook
results/retrieval_metrics_comparison.csv	Five-system retrieval summary
results/retrieval_detail_comparison.csv	Per-question held-out retrieval results
results/generation_eval_questions_comparison.csv	Fixed 30-question generation subset
results/generation_eval_outputs_comparison.csv	All 150 generated outputs (30 × 5 systems)
results/generation_metrics_comparison.csv	Aggregated generation metrics
results/adaptive_weight_search_comparison.csv	TRAIN-only search for Adaptive fusion strength
results/router_*.csv	Router OOF training and DEV diagnostics
results/abstention_*.csv	Evidence-sufficiency calibration and evaluation
results/adaptive_vs_*.csv	Adaptive-vs-baseline pairwise generation comparisons
results/experiment_config_comparison.json	Experiment settings
manifest.json	SHA-256 checksum and file size for released files

---

11. Quick inspection without rerunning the notebook

Clone the repository or download the CSV files and inspect them directly with pandas:

import pandas as pd

retrieval = pd.read_csv("results/retrieval_metrics_comparison.csv")
generation = pd.read_csv("results/generation_metrics_comparison.csv")
outputs = pd.read_csv("results/generation_eval_outputs_comparison.csv")

print(retrieval)
print(generation)

# Compare all five answers for one evaluation question.
qid = outputs["id"].iloc[0]

print(
    outputs.loc[
        outputs["id"].eq(qid),
        ["system", "question", "reference_answer", "generated_answer"]
    ].to_string(index=False)
)

---

12. Reproducing the experiment

The easiest setup is Google Colab with a T4-class GPU.

1. Open TechQA_Hybrid_RAG_Qwen25_7B_System_Comparison.ipynb.
2. Select a GPU runtime.
3. Run the notebook from top to bottom.
4. The notebook downloads the public upstream dataset and pretrained models.
5. It rebuilds the reduced corpus, chunks, embeddings, and indexes.
6. It trains/selects the lightweight Adaptive router using TRAIN only.
7. It evaluates all five retrieval strategies on answerable DEV.
8. It evaluates evidence-sufficiency on DEV.
9. It runs the fixed 30-question generation comparison.
10. It exports the detailed CSV/JSON results.

The generation stage is the slowest stage because Qwen is run for every system/question pair.

Main dependencies

• datasets
• transformers
• accelerate
• bitsandbytes
• sentence-transformers
• faiss-cpu
• rank-bm25
• rouge-score
• pandas
• scikit-learn

See requirements.txt for the release environment.

---

13. Intended use

This repository is intended for:

• RAG retrieval research;
• technical-support retrieval evaluation;
• BM25 vs dense vs hybrid comparison;
• query-adaptive fusion experiments;
• reranking ablation studies;
• analysis of retrieval-vs-generation behavior;
• reproducibility and educational use.

It can also serve as a reference implementation for building lightweight query-adaptive hybrid retrievers without fine-tuning the underlying embedding model.

---

14. Limitations

Please keep the following limitations in mind:

1. Reduced corpus
   The experiment uses the reduced TechQA-RAG-Eval corpus, not the original full TechQA / full IBM Technotes collection.

2. Generation subset size
   End-to-end generation is evaluated on 30 answerable DEV questions rather than all 160 answerable DEV questions because generation is substantially more expensive.

3. No significance claim
   Differences between systems should not be interpreted as statistically significant unless paired significance testing is added.

4. Automated generation metrics are incomplete
   ROUGE-L and embedding similarity do not fully capture technical correctness, completeness, or harmful inaccuracies.

5. Frozen generator
   Qwen is not fine-tuned for this task. Some failures arise after successful retrieval because the generator may abstain or under-use evidence.

6. Router class imbalance
   Most TRAIN router targets are Balanced; BM25-heavy examples are relatively rare. Balanced accuracy and macro F1 should therefore be considered alongside raw router accuracy.

7. Reranker conclusion is experiment-specific
   The neural reranker underperforms in this controlled setup. This should not be generalized to all rerankers, datasets, or candidate-selection strategies.

8. Technical-support outputs require verification
   Generated answers are experimental outputs and should not be treated as authoritative production support guidance without checking the cited source documentation.

---

15. Dataset provenance and license

The upstream dataset is nvidia/TechQA-RAG-Eval, a reduced version of the original TechQA benchmark designed for RAG evaluation.

The upstream Hugging Face dataset card lists the license as Apache-2.0.

This repository does not republish the full upstream source corpus as a new dataset. Source-derived cache files are excluded from the public release by default and can be rebuilt through the notebook.

Original TechQA reference:

• IBM TechQA: https://github.com/IBM/techqa

---

16. Reproducibility safeguards

The experiment is designed to avoid DEV leakage:

• router targets are derived from TRAIN;
• router model selection uses TRAIN out-of-fold predictions;
• Adaptive fusion strength is selected on TRAIN;
• abstention threshold is selected on TRAIN out-of-fold predictions;
• DEV references are used only for held-out evaluation;
• the same fixed generation question set is used for all five systems;
• the same generator and evidence budget are used for all systems.

---

17. Main conclusion

The controlled comparison supports the following conclusion:

Adaptive Hybrid RRF achieved the strongest overall held-out document-ranking performance and was therefore selected as the primary retrieval strategy. Static Hybrid RRF remained highly competitive and achieved the strongest automated generation metrics on the fixed 30-question generation subset. Dense BGE remained strongest at deeper recall, while the evaluated neural reranker did not provide a reliable improvement.

The main finding is therefore not that one system wins every metric.

Instead:

Query-adaptive BM25/Dense fusion can improve overall top-rank retrieval quality, while downstream answer quality remains sensitive to passage selection, evidence budgeting, abstention behavior, and the frozen generator.

---

18. Associated selected system

A separate Hugging Face Model repository will package the selected Adaptive Hybrid RRF retrieval system for easier reuse.

Once published, its link will be added here.

The comparison repository remains the source of the experimental evidence used to justify that selection.

---

Citation

If you use this experiment package, please cite the repository:

@misc{vltruong01_techqa_hybrid_rag_2026,
  author       = {vltruong01},
  title        = {TechQA Hybrid RAG System Comparison},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/datasets/vltruong01/TechQA-Hybrid-RAG-System-Comparison}
}

Please also follow the citation / attribution guidance of the upstream TechQA and TechQA-RAG-Eval resources.

---

Upstream models and resources

• Dataset: nvidia/TechQA-RAG-Eval
• Original TechQA: IBM/techqa
• Dense retriever: BAAI/bge-base-en-v1.5
• Neural reranker: BAAI/bge-reranker-base
• Generator: Qwen/Qwen2.5-7B-Instruct

---

Repository

Hugging Face Dataset:
https://huggingface.co/datasets/vltruong01/TechQA-Hybrid-RAG-System-Comparison