RAG Search Engine: Full Technical Documentation

This repository implements a movie search system that combines lexical retrieval, dense semantic retrieval, reciprocal-rank fusion, optional reranking, multimodal query support, and retrieval-augmented generation.

The codebase is organized as runnable CLIs over shared retrieval libraries, with cached indexes/embeddings and optional Gemini-based query enhancement and judging.

Youtube Video

Youtube Link

1. System Scope and What Is Actually Implemented

Implemented:

• Keyword retrieval with inverted index + BM25: cli/lib/keyword_search.py
• Semantic retrieval with sentence embeddings + cosine similarity: cli/lib/semantic_search.py
• Chunk-based semantic retrieval (max-score aggregation per document): cli/lib/semantic_search.py
• Hybrid retrieval:
  • Weighted score fusion API: HybridSearch.weighted_search
  • Reciprocal Rank Fusion (RRF): HybridSearch.rrf_search
  • cli/hybrid_search_cli.py
• Query enhancement with Gemini (spell, rewrite, expand): utils/hybrid_search_utils/query_enhancement.py
• Reranking:
  • Gemini individual scoring
  • Gemini batch list ranking
  • Local cross-encoder reranking
  • utils/hybrid_search_utils/rerank_methods.py
• Evaluation:
  • Golden-set precision/recall/F1@k: cli/evaluation_cli.py
  • LLM-as-a-judge 0-3 relevance: utils/hybrid_search_utils/evaluate.py
• Multimodal search (image query against text corpus using CLIP encoder): cli/lib/multimodal_search.py
• Image-informed query rewrite (Gemini vision): utils/image_cli_utils/gemini.py, cli/describe_image_cli.py
• Retrieval-Augmented Generation workflows (rag, summarize, citations, question): cli/augmented_generation_cli.py, utils/augmented_utils/gemini.py

Not implemented (important):

• No vector database (FAISS/Chroma/Pinecone/Milvus/Qdrant are not used)
• No ANN index; semantic search is brute-force cosine over all vectors
• No learned hybrid weighting or feedback loop training

2. High-Level Architecture

2.1 End-to-end query pipeline (RRF + optional rerank + optional generation)

User Query
|
+--> Optional Query Enhancement (spell/rewrite/expand via Gemini)
|
+--> BM25 Search (Inverted Index)
|
+--> Semantic Search (Chunk Embeddings + Cosine)
|
+--> Fusion
|- Weighted Fusion (alpha blend) OR
\- Reciprocal Rank Fusion (RRF)
|
+--> Optional Reranking
|- Gemini individual score
|- Gemini batch ordering
\- Cross-Encoder (ms-marco-TinyBERT-L2-v2)
|
+--> Optional LLM Evaluation (0-3 judge)
|
+--> Optional Augmented Generation (Gemini answer/summarize/citations)
|
Final Ranked Results / Generated Response

2.2 Indexing and cache pipeline

movies.json
|
+--> Keyword indexing
| - tokenization / stopword removal / stemming
| - inverted postings + TF + doc lengths
| - cache/index.pkl, docmap.pkl, term_frequency.pkl, doc_length.pkl
|
+--> Semantic indexing
- text normalization
- bi-encoder embeddings (all-MiniLM-L6-v2)
- chunk embeddings + metadata
- cache/movie_embeddings.npy
- cache/chunk_embeddings.npy
- cache/chunk_metadata.json

2.3 Multimodal branch

Input Image
|
+--> CLIP image embedding (clip-ViT-B-32)
|
+--> Cosine similarity vs CLIP text embeddings of movie title+description
|
Top-K movies (image-text retrieval)

3. Data and Runtime Components

• Dataset: data/movies.json (5000 movie documents)
• Golden evaluation set: data/golden_dataset.json (10 test queries)
• Stopword file: data/stopwords.txt
• Cache directory (auto-generated): cache/
• Environment variable for Gemini features: GEMINI_API_KEY

4. Module-by-Module Implementation Details

4.1 Keyword search internals

Primary implementation: cli/lib/keyword_search.py

Core structures:

• index: dict[str, list[int]]: token -> postings list of doc IDs
• docmap: dict[int, dict]: doc ID -> full movie record
• term_frequency: dict[int, Counter]: per-document term counts
• doc_length: dict[int, int]: tokenized doc lengths

Build path:

• build(movies) calls private __add_document(doc_id, text)
• Indexed text is "{title} {description}"
• Preprocessing uses utils/keyword_seach_utils/text_preprocessing.py

Search path:

• bm25_search(query, limit) scores every document for all query tokens
• Returns top-k (doc_id, score) sorted descending

Formulas implemented:

• Raw TF: tf(d,t) from term_frequency
• IDF: idf(t) = ln((N + 1) / (df(t) + 1))
• BM25 TF saturation: bm25_tf(d,t) = tf(d,t) * (k1 + 1) / (tf(d,t) + k1 * (1 - b + b * |d|/avgdl))
• BM25 IDF: bm25_idf(t) = ln(((N - df(t) + 0.5)/(df(t) + 0.5)) + 1)
• BM25 term score: bm25(d,t) = bm25_tf(d,t) * bm25_idf(t)

Defaults:

• k1 = 1.5, b = 0.75 from utils/keyword_seach_utils/search_utils.py

Caching:

• save() writes 4 pickle files into cache/
• load(documents) loads cache or rebuilds if missing

CLI surface:

• cli/keyword_search_cli.py
  • build, search, tf, idf, bm25idf, tfidf, bm25tf, bm25search

4.2 Semantic search internals

Primary implementation: cli/lib/semantic_search.py

Bi-encoder model:

• SentenceTransformer("all-MiniLM-L6-v2")

Document embedding:

• Uses string form: "{title}: {description}"
• Stored in cache/movie_embeddings.npy

Query embedding:

• generate_embedding(text) with semantic text preprocessing:
  • unicode normalization, lowercase, punctuation strip, whitespace cleanup
  • utils/semantic_search_utils/text_preprocessing.py

Similarity:

• cosine_similarity(query_vec, doc_vec) in utils/semantic_search_utils/vector_operations.py
• Brute-force over all embeddings; sorted descending

Chunked semantic retrieval (ChunkedSemanticSearch):

• Splits each description via semantic_chunk(text, 4, 1)
• Encodes all chunks, stores:
  • cache/chunk_embeddings.npy
  • cache/chunk_metadata.json (movie_idx/chunk_idx/total_chunks)
• search_chunk(query, limit):
  • computes query-chunk cosine
  • keeps max chunk score per movie
  • ranks movies by that max score

Why chunking helps here:

• Long descriptions can dilute full-document embeddings
• Max-over-chunks better preserves local relevance signals

4.3 Hybrid retrieval internals

Primary implementation: cli/lib/hybrid_search.py

HybridSearch initialization:

• Loads chunk semantic index (ChunkedSemanticSearch)
• Loads keyword index (InvertedIndex)
• Maintains document_map

Weighted fusion

Path: HybridSearch.weighted_search(query, alpha, limit)

Behavior:

• Runs BM25 and semantic searches with expanded candidate size limit*50
• Min-max normalizes each score list via utils/hybrid_search_utils/normalize_score.py
• Combines per doc via weighted sum

Formula (intended): hybrid_score = alpha * bm25_norm + (1 - alpha) * semantic_norm

Helper:

• utils/hybrid_search_utils/score_utils.py::hybrid_score

Reciprocal Rank Fusion (RRF)

Path: HybridSearch.rrf_search(query, k, limit)

Behavior:

• Gets BM25 and semantic ranked lists (limit*50 candidates internally)
• Assigns ranks per list
• Fuses by reciprocal rank contribution from each retriever

Formula: RRF(d) = 1/(k + rank_bm25(d)) + 1/(k + rank_semantic(d))

• Missing rank contributes 0
• k defaults to 60 in CLI

Helper:

• utils/hybrid_search_utils/score_utils.py::rrf_score

CLI surface:

• cli/hybrid_search_cli.py
  • weighted_search
  • rrf_search / alias rrf-search
  • options: --enhance, --rerank-method, --debug, --evaluate, --json

4.4 Query enhancement

Implementation: utils/hybrid_search_utils/query_enhancement.py

Methods:

• spell: typo correction
• rewrite: concise specific rewrite using movie priors
• expand: append related terms/synonyms

All methods call Gemini (gemini-2.5-flash) and return text used as modified query.

Design impact:

• Helps lexical recall and may improve semantic focus
• Adds API latency/cost and introduces model variance

4.5 Reranking layer

Implementation: utils/hybrid_search_utils/rerank_methods.py

Modes:

• individual:
  • prompt each result independently for score 0-10
  • parse float via regex fallback
  • sleeps 3 seconds between calls
• batch:
  • prompt with whole candidate list
  • ask for JSON ordered IDs
  • parse and reorder by returned ID ranking
• cross_encoder:
  • local model cross-encoder/ms-marco-TinyBERT-L2-v2
  • score query-document pairs jointly

API entrypoint:

• rerank(results, rerank_method, query, documents, limit)

Output:

• Annotates scores (rerank_score or cross_encoder_score)
• Returns top limit

4.6 Evaluation layer

Automatic relevance metrics (golden set)

Implementation: cli/evaluation_cli.py

Process:

1. Load data/golden_dataset.json
2. For each query, run subprocess:
   • uv run cli/hybrid_search_cli.py rrf_search <query> --k=60 --limit=<k> --json
3. Compare retrieved titles to known relevant titles
4. Compute metrics per query

Formulas:

• Precision@k = (# relevant retrieved in top k) / k
• Recall@k = (# relevant retrieved in top k) / (# relevant documents)
• F1@k = 2 * P@k * R@k / (P@k + R@k) (0 if denominator is 0)

LLM-as-a-judge evaluation

Implementation: utils/hybrid_search_utils/evaluate.py

Process:

• Prompts Gemini to rate each result on 0..3 scale
• Requires strict JSON array output length == number of results
• Parsing is robust to fenced/prose output

Scale:

• 3 highly relevant
• 2 relevant
• 1 marginally relevant
• 0 not relevant

4.7 Augmented generation layer (RAG)

CLIs and helpers:

• cli/augmented_generation_cli.py
• utils/augmented_utils/rrf_search.py
• utils/augmented_utils/gemini.py

Retrieval path:

• Always retrieves via hybrid RRF (k=60) using subprocess call to cli/hybrid_search_cli.py ... --json

Generation modes:

• rag: general answer generation
• summarize: condensed synthesis across retrieved movies
• citations: source-citation style answer ([1], [2], ... requested)
• question: conversational QA style response

Prompting strategy:

• retrieved docs are injected directly as serialized Python list content
• no token-aware truncation/chunk selection policy implemented
• no explicit groundedness verifier/citation checker implemented

4.8 Multimodal retrieval layer

Implementation:

• cli/lib/multimodal_search.py
• cli/multimodal_search_cli.py

Model:

• SentenceTransformer("clip-ViT-B-32")

Indexing:

• Encodes all text documents ("title: description") in CLIP text space

Query:

• Encodes input image via PIL + CLIP image encoder
• Computes cosine similarity vs text embeddings
• Returns top 5 movies

Image-assisted query rewriting:

• utils/image_cli_utils/gemini.py
• cli/describe_image_cli.py
• Gemini receives image bytes + text query and rewrites search query

5. Concept Glossary (Definitions, Formula, Intuition, Usage, Pros/Cons)

Each concept below is tied to repository code.

Term Frequency (TF)

• Definition: number of times term t appears in document d.
• Formula: TF(d,t) = count(d,t)
• Intuition: repeated words can indicate topical importance.
• Used in: cli/lib/keyword_search.py::get_tf
• Pros: simple, interpretable.
• Cons: can over-reward long/repetitive documents.
• Use when: exact lexical matching matters.

Inverse Document Frequency (IDF)

• Definition: rarity weighting for a term across corpus.
• Formula (implemented): IDF(t) = ln((N+1)/(df(t)+1))
• Intuition: rare terms carry more discriminative power.
• Used in: cli/lib/keyword_search.py::get_idf
• Pros: downweights ubiquitous terms.
• Cons: unstable on very small corpora.
• Use when: scoring should penalize common words.

TF-IDF

• Definition: product of term frequency and inverse document frequency.
• Formula: TFIDF(d,t) = TF(d,t) * IDF(t)
• Intuition: high if term is frequent in doc and rare globally.
• Used in: cli/keyword_search_cli.py::tfidf command.
• Pros: baseline sparse ranking feature.
• Cons: no length saturation/normalization by default.
• Use when: lightweight lexical scoring/debugging.

BM25

• Definition: probabilistic lexical ranking with TF saturation and length normalization.
• Formulas used:
  • bm25_tf = tf*(k1+1)/(tf + k1*(1-b + b*|d|/avgdl))
  • bm25_idf = ln(((N-df+0.5)/(df+0.5))+1)
  • bm25 = bm25_tf * bm25_idf
• Used in: cli/lib/keyword_search.py
• Pros: strong classical baseline, robust lexical retrieval.
• Cons: cannot capture semantic paraphrase.
• Use when: exact tokens and explainable ranking are needed.

BM25-IDF

• Definition: the BM25-specific IDF term above.
• Used in: get_bm25_idf and bm25idf CLI command.
• Note: differs from plain IDF smoothing.

Semantic Search

• Definition: retrieve by vector similarity in embedding space.
• Formula: cosine similarity between query/document vectors.
• Used in: cli/lib/semantic_search.py
• Pros: handles synonymy and paraphrase.
• Cons: less transparent, model-dependent.
• Use when: meaning match > exact wording.

Embeddings

• Definition: dense vectors encoding semantic content.
• Used models:
  • all-MiniLM-L6-v2 (text)
  • clip-ViT-B-32 (text-image shared space)
• Used in: cli/lib/semantic_search.py, cli/lib/multimodal_search.py

Vector Space Model

• Definition: represent text/image as vectors and compute geometric similarity.
• Used in: semantic and multimodal retrieval modules.

Cosine Similarity

• Definition: angle-based similarity between vectors.
• Formula: cos(a,b) = (a·b)/(||a|| ||b||)
• Used in: utils/semantic_search_utils/vector_operations.py
• Pros: scale-invariant.
• Cons: still O(N) scan in this codebase.

Hybrid Search

• Definition: combine lexical and semantic retrievers.
• Implemented strategies:
  • weighted normalized score fusion
  • RRF rank fusion
• Used in: cli/lib/hybrid_search.py

Weighted Scoring

• Definition: linear blend of normalized retriever scores.
• Formula: alpha*sparse + (1-alpha)*dense
• Used in helper: utils/hybrid_search_utils/score_utils.py::hybrid_score
• Risk: sensitive to calibration and normalization quality.

RRF (Reciprocal Rank Fusion)

• Definition: fuse based on rank positions, not raw scores.
• Formula: sum_i 1/(k + rank_i(d))
• Used in: cli/lib/hybrid_search.py::rrf_search
• Pros: robust across heterogeneous score scales.
• Cons: ignores absolute score margins.

Cross-Encoder

• Definition: reranker that scores (query, doc) jointly in one model forward pass.
• Used in: rerank_cross_encoder with ms-marco-TinyBERT-L2-v2.
• Pros: usually better precision at top ranks.
• Cons: higher latency than bi-encoder retrieval.

Bi-Encoder

• Definition: independently encode query and docs, compare vectors.
• Used in: SemanticSearch and ChunkedSemanticSearch.
• Pros: scalable precompute + fast similarity.
• Cons: less precise than cross-encoder rerank.

Reranking

• Definition: reorder top candidates after first-pass retrieval.
• Used in: Gemini (individual, batch) and cross-encoder methods.

Query Expansion

• Definition: append related terms to broaden lexical match.
• Used in: enhance_query(..., "expand").
• Benefit: recall gains.
• Risk: topic drift.

Query Enhancement / Rewrite

• Definition: model-driven reformulation for specificity/clarity.
• Used in: spell, rewrite, expand; and image-conditioned rewrite path.

Precision / Recall / F1

• Precision@k: relevance purity of top-k.
• Recall@k: coverage of relevant set in top-k.
• F1@k: harmonic mean balancing both.
• Used in: cli/evaluation_cli.py

Manual Evaluation

• Definition: human judging relevance quality.
• Status in repo: golden dataset appears manually curated (data/golden_dataset.json).

LLM-based Evaluation

• Definition: LLM judges each result on a rubric.
• Used in: utils/hybrid_search_utils/evaluate.py
• Risk: judge bias/variance and prompt sensitivity.

Multimodal Search

• Definition: retrieval across modalities (image query to text docs in shared embedding space).
• Used in: cli/lib/multimodal_search.py

Retrieval-Augmented Generation (RAG) / Augmented Generation

• Definition: retrieve supporting docs, then generate answer conditioned on them.
• Used in: cli/augmented_generation_cli.py + Gemini helpers.
• Difference from retrieval-only: adds synthesis/generation, not just ranking.

Relevance Scoring

• Implemented forms:
  • BM25 score
  • Cosine similarity
  • RRF score
  • Cross-encoder score
  • Gemini rerank score

Embedding Indexing

• Definition: precompute/store corpus embeddings for later query-time similarity.
• Used in: .npy caches for semantic and chunk embeddings.

Vector Database

• Status: not used in this project.
• Current approach: in-memory numpy arrays + linear scan.

ANN (Approximate Nearest Neighbors)

• Status: not implemented.
• Implication: exact scan with O(N) query-time complexity in vector retrieval.

6. Design Decisions and Alternatives

Why not pure keyword search?

• Pure BM25 fails on synonymy/paraphrases and indirect phrasing.
• Semantic branch addresses this by matching meaning, not just tokens.

Why not pure semantic search?

• Dense retrieval can miss exact rare entities/spellings; lexical precision for named terms can be stronger.
• BM25 branch preserves hard lexical constraints.

Why hybrid?

• Combines lexical precision and semantic recall.
• More robust across varied query types in movie discovery.

Why RRF over weighted-only fusion?

• Weighted fusion depends on score calibration and normalization quality.
• RRF uses ranks, typically more stable across retrievers with different score distributions.

Why cross-encoder reranking?

• First-pass retrieval is recall-oriented.
• Cross-encoder improves top-k precision by modeling fine query-document interactions.

Why Gemini-based reranking/evaluation?

• Enables semantic judgments without building separate supervised reranker/judge datasets.
• Tradeoff: API latency/cost and non-determinism.

Why multimodal extension?

• Movie discovery often starts from visual cues (poster/frame/mood).
• CLIP-style shared space enables image-to-text retrieval.

7. Comparison Tables

7.1 Keyword vs Semantic Search

Feature	Keyword (BM25)	Semantic (Embeddings)
Matching basis	token overlap	vector similarity
Synonym handling	weak	strong
Exact entity matching	strong	moderate
Explainability	high	medium/low
Latency profile	low (after index)	medium (embedding + scan)
Typical failure mode	wording mismatch	semantic drift / missed exact token
Best use case	precise known-item queries	natural language intent queries

7.2 Weighted Hybrid vs RRF

Feature	Weighted Fusion	RRF
Input	normalized scores	ranks
Calibration sensitivity	high	low
Interpretability	medium	high (rank-based)
Robustness across retrievers	medium	high
Parameterization	alpha	k
In this repo	present but currently buggy call path	primary production path

7.3 Cross-Encoder vs Bi-Encoder

Feature	Bi-Encoder	Cross-Encoder
Encode docs offline	yes	no
Retrieval speed	fast	slower
Pairwise interaction modeling	no	yes
Top-k precision	moderate	higher
Role here	first-pass retrieval	reranker

7.4 Manual vs LLM Evaluation

Feature	Manual Labels	LLM Judge
Consistency	depends on rubric/humans	prompt/model dependent
Cost at scale	high	lower marginal
Transparency	high	medium
Bias risk	annotator bias	model bias/prompt bias
In this repo	golden set relevance lists	optional 0-3 CLI evaluation

8. Evaluation Methodology in This Repository

8.1 Golden-set metric evaluation

• Ground truth source: data/golden_dataset.json
• Query set size: 10
• For each query:
  • Run RRF hybrid retrieval
  • Compare top-k titles to list of relevant titles
  • Print precision@k, recall@k, F1@k

Strengths:

• deterministic metric computation
• easy baseline comparison

Limitations:

• small benchmark size
• exact title matching only
• no graded relevance beyond binary relevant/not relevant

8.2 LLM-as-a-judge evaluation

• Optional from hybrid CLI: --evaluate
• Judge rubric is 0..3 relevance scale
• Parsing enforces strict integer JSON output

Strengths:

• captures nuanced semantic relevance
• faster than full human annotation loop

Risks and bias:

• model may be inconsistent across runs
• prompt wording can alter scores
• possible over-leniency or over-strictness by domain/query type

Recommendation:

• Use LLM judging as supplemental signal, not sole acceptance criterion.

9. Multimodal Pipeline Documentation

Implemented multimodal capabilities are twofold:

1. Image-to-movie retrieval (cli/multimodal_search_cli.py image_search)

• Build CLIP text embeddings for all movie documents
• Encode query image
• Compute image-text cosine similarity
• Return top matches

2. Image-conditioned query rewriting (cli/describe_image_cli.py)

• Send image + text query to Gemini
• Return rewritten query for downstream text retrieval

Difference vs pure text pipeline:

• Text-only path encodes only language with MiniLM/lexical methods.
• Multimodal path uses CLIP shared embedding space and can use visual evidence directly.

10. Augmented Generation Documentation

What augmented generation means here

• Retrieval and generation are separate stages.
• Retrieval finds candidate supporting documents.
• Generation synthesizes final answer from retrieved context.

Implemented flow

• Retrieval: rrf_search JSON results via subprocess
• Generation: Gemini prompt with serialized results
• Modes: answer/summarize/citations/conversational QA

Prompt construction

• Prompts pass query/question + raw docs list as context.
• The system asks for concise, user-facing output tailored to Hoopla users.

Context-window considerations

• Current implementation does not explicitly trim by token budget.
• Risk grows with larger --limit and long descriptions.

Hallucination mitigation status

Implemented:

• retrieval grounding by passing documents
• citation-mode prompting

Not implemented:

• hard citation verification
• abstention classifier beyond prompt instructions
• source-faithfulness checker

11. Implementation Reference Map (File-Level)

Core retrieval:

• cli/lib/keyword_search.py: inverted index, TF/IDF/BM25, cache IO
• utils/keyword_seach_utils/text_preprocessing.py: lexical preprocessing (stopwords + Porter stemming)
• utils/keyword_seach_utils/search_utils.py: BM25 constants (k1, b)
• cli/lib/semantic_search.py: sentence-transformer retrieval + chunked retrieval
• utils/semantic_search_utils/semantic_chunk.py: sentence-window chunking
• utils/semantic_search_utils/vector_operations.py: cosine similarity and vector ops
• utils/semantic_search_utils/text_preprocessing.py: normalization for embedding inputs

Hybrid and fusion:

• cli/lib/hybrid_search.py: orchestrates BM25 + chunk-semantic and fusion
• utils/hybrid_search_utils/score_utils.py: weighted + RRF formula helpers
• utils/hybrid_search_utils/normalize_score.py: min-max normalization

Rerank and query enhancement:

• utils/hybrid_search_utils/query_enhancement.py: spell/rewrite/expand via Gemini
• utils/hybrid_search_utils/rerank_methods.py: Gemini rerank and cross-encoder rerank
• utils/hybrid_search_utils/print.py: result formatting
• utils/hybrid_search_utils/debugger.py: debug logging helpers
• utils/hybrid_search_utils/evaluate.py: LLM relevance scoring helper

Evaluation and generation CLIs:

• cli/evaluation_cli.py: P/R/F1 evaluation against golden set
• cli/hybrid_search_cli.py: main hybrid CLI with enhancement/rerank/eval flags
• cli/augmented_generation_cli.py: retrieval + generation workflows
• utils/augmented_utils/rrf_search.py: robust subprocess JSON parse wrapper
• utils/augmented_utils/gemini.py: generation/summarization/citation prompts

Multimodal:

• cli/lib/multimodal_search.py: CLIP image-text retrieval
• cli/multimodal_search_cli.py: multimodal CLI entrypoint
• utils/image_cli_utils/gemini.py: image-conditioned query rewrite
• cli/describe_image_cli.py: command wrapper for image rewrite

Data/loading:

• utils/cli_utils/file_loading.py: dataset loading (movies.json)
• data/movies.json: corpus
• data/golden_dataset.json: evaluation cases
• data/stopwords.txt: keyword preprocessing stopwords

12. Known Limitations and Current Gaps

These are code-level issues visible in the current repository state:

• cli/keyword_search_cli.py imports utils.keyword_seach_utils.load_utils, but this module is missing.
• utils/keyword_seach_utils/tfidf_utils.py calls InvertedIndex.load() without required documents argument.
• cli/hybrid_search_cli.py has normalize command path calling normalize_score without importing it.
• cli/lib/hybrid_search.py::weighted_search calls hybrid_score without passing alpha argument.
• Semantic retrieval assumes mapping by i+1 in SemanticSearch.search, which can break if document IDs are not contiguous from 1.
• No ANN/vector DB support; dense retrieval is linear scan and may not scale for large corpora.
• No distributed indexing or sharding.
• No online learning from clicks/feedback.
• RAG context assembly is naive (full serialized docs, no token-budget optimizer).
• LLM-dependent modules require network/API key and are non-deterministic.

13. Scalability, Latency, and Memory Tradeoffs

• Sparse BM25 index in Python dict/list is memory-heavy but straightforward.
• Dense retrieval stores full embeddings in memory/NumPy arrays; query-time complexity is O(N*d).
• Chunked semantic search improves recall granularity but increases index size and compute.
• Cross-encoder reranking improves precision but increases latency roughly O(k) model inferences.
• Gemini-based enhancement/rerank/eval adds network latency and cost per request.

15. CLI Entry Points Summary

• Keyword: cli/keyword_search_cli.py
• Semantic: cli/semantic_search_cli.py
• Hybrid: cli/hybrid_search_cli.py
• Evaluation: cli/evaluation_cli.py
• Multimodal: cli/multimodal_search_cli.py
• Image-assisted rewrite: cli/describe_image_cli.py
• Augmented generation: cli/augmented_generation_cli.py

16. Environment and Dependencies

Defined in pyproject.toml:

• sentence-transformers
• numpy
• nltk
• google-genai
• python-dotenv
• pillow

Python requirement:

• >=3.13

For Gemini features (query_enhancement, rerank with Gemini, LLM evaluation, RAG generation, image rewrite):

• set GEMINI_API_KEY in environment or .env

---

If you are extending this system, start in cli/lib/hybrid_search.py and utils/hybrid_search_utils/rerank_methods.py: these are the main orchestration points where retrieval quality/latency tradeoffs are controlled.