README.md

metadata

title: VECTOR VisionQuest
emoji: 🚀
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 5.20.0
app_file: main.py
pinned: false

⚡ VECTOR: Voice-Enabled Multilingual Indic RAG Engine

An instrumented, ultra-low-latency, voice-enabled Retrieval-Augmented Generation (RAG) engine built from scratch for 15 Indic languages.

🌐 Live Deployed Application: https://hungry-games-dance.loca.lt

---

📌 Executive Summary

VECTOR is an open-source, high-throughput, sub-10ms Retrieval-Augmented Generation (RAG) engine engineered specifically for the linguistic diversity of the Indian subcontinent. Operating on low-cost CPU environments, VECTOR delivers end-to-end voice and text question answering across 14 Indic languages (Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Sanskrit, Tamil, Telugu, Urdu) plus English (15 languages total, ~743,000 deduplicated passages).

The active runtime deployment loads 148,854 in-memory FAISS vectors across 3 core active languages (English [en], Hindi [hi], and Marathi [mr]), achieving an average retrieval latency of ~7.04 ms (p95: 7.97 ms vs 50.0 ms budget SLA).

Key Architectural Advantages

• ⚡ Sub-10ms Vector Retrieval: In-memory FAISS HNSW graph traversal ($0.73\text{ ms}$) + INT8 ONNX vectorized embedding ($6.31\text{ ms}$) on CPU.
• 🛡️ Cascaded 4-Tier Guardrails: Stem regex with variable word-gap sliding, Meta Prompt-Guard 86M neural DPI/IPI shield, 6-class intent filter, and own-language centroid distance gate.
• 🔀 Script-Aware BM25 + Dense Fusion: Automatic cross-script detection bypassing lexical penalties for cross-lingual queries.
• 🧮 Deterministic Context Synthesis: TextRank graph centrality + SVD singular energy matrix reduction delivering factual answers in $<10\text{ ms}$ on CPU with zero LLM API cost or latency.
• 🌴 Command Center UI: Retro-tropical Web Audio frequency visualizer with real-time 9-stage telemetry waterfall breakdown.

---

🏛️ Architecture

graph LR
    subgraph PATH1 ["🎙️ Path 1: Audio & Text Ingestion"]
        A[Audio Upload / Microphone Stream] --> STT[Sarvam Saaras STT + ffmpeg 16kHz]
        T[Raw Text Input Bypass] --> ROUTER[Language Resolution Router]
        STT --> ROUTER
    end

    subgraph PATH2 ["🛡️ Path 2: 4-Tier Security Shield"]
        ROUTER --> G1[Tier-1 Stem Regex + Obfuscation Decoder]
        G1 -- Safe --> G2[Tier-2 Meta Prompt-Guard 86M DPI]
        G2 -- Safe --> G3[Tier-3 6-Class Query Intent Gate]
        G3 -- Factual --> G4[Tier-4 Own-Lang Centroid Distance Gate]
    end

    subgraph PATH3 ["⚡ Path 3: Sub-0.5ms Hot Cache Fast-Path"]
        G4 -- On-Topic --> CACHE{Hot Cache Lookup}
        CACHE -- "Hit (<0.5ms)" --> FAST_OUT[Zero-Latency Response]
    end

    subgraph PATH4 ["🔎 Path 4: Hybrid Vector Retrieval Engine"]
        CACHE -- "Miss" --> EMB[multilingual-e5-small INT8 ONNX]
        EMB --> FAISS[Parallel FAISS HNSW Native & LongDoc Search]
        FAISS --> RRF[Reciprocal Rank Fusion k=60]
        RRF --> BM25[Adaptive Script-Aware BM25 Fusion]
        BM25 --> GATE{Disqualification Gate}
    end

    subgraph PATH5 ["🧠 Path 5: Deterministic Synthesis & Grounding"]
        GATE -- High Relevance --> IPI[Batched Prompt-Guard IPI Context Scan]
        IPI -- Clean Chunks --> SYNTH[Continuous TextRank + SVD Energy Synthesis]
        SYNTH --> GROUND[Post-Gen Grounding Overlap Verifier]
        GROUND -- Grounded --> FINAL_OUT[JSON Response + 9-Stage Telemetry]
    end

    %% Rejection Routing
    G1 -- Blocked --> REJECT[Declined Response: Safety Violation]
    G2 -- Injected --> REJECT
    G3 -- Non-Factual --> REJECT
    G4 -- Off-Topic --> REJECT
    GATE -- Score < 0.35 --> DECLINE[Declined Response: Insufficient Info]
    IPI -- Poisoned --> REJECT
    GROUND -- Ungrounded --> DECLINE

    %% Custom Styling Classes
    classDef inputStyle fill:#1E293B,stroke:#38BDF8,stroke-width:2px,color:#F8FAFC;
    classDef sttStyle fill:#0F172A,stroke:#818CF8,stroke-width:2px,color:#F8FAFC;
    classDef guardStyle fill:#311B92,stroke:#B388FF,stroke-width:2px,color:#FFFFFF;
    classDef cacheStyle fill:#064E3B,stroke:#34D399,stroke-width:2px,color:#F8FAFC;
    classDef faissStyle fill:#164E63,stroke:#22D3EE,stroke-width:2px,color:#F8FAFC;
    classDef synthStyle fill:#4C1D95,stroke:#C084FC,stroke-width:2px,color:#F8FAFC;
    classDef outStyle fill:#065F46,stroke:#10B981,stroke-width:2px,color:#FFFFFF;
    classDef blockStyle fill:#881337,stroke:#F43F5E,stroke-width:2px,color:#FFFFFF;

    class A,T inputStyle;
    class STT,ROUTER sttStyle;
    class G1,G2,G3,G4 guardStyle;
    class CACHE cacheStyle;
    class EMB,FAISS,RRF,BM25,GATE faissStyle;
    class IPI,SYNTH,GROUND synthStyle;
    class FAST_OUT,FINAL_OUT outStyle;
    class REJECT,DECLINE blockStyle;

🛣️ Swimlane Pipeline Execution Breakdown

Swimlane / Path	Key Components & Models	Latency Budget	Action on Failure / Edge Case
🎙️ Path 1: Ingestion & STT	Sarvam Saaras saaras:v3 + ffmpeg 16kHz mono normalizer	< 150 ms (Audio) / < 0.1 ms (Text)	Fallback to default language_hint or auto-detect
🛡️ Path 2: 4-Tier Security Shield	Tier-1 Regex Stem Gap=4, Tier-2 Prompt-Guard 86M ONNX, Tier-3 6-Class Intent, Tier-4 Centroid Distance	< 2.5 ms	Fail-Safe-by-Category (model_failed=True), block immediately
⚡ Path 3: Cache Fast-Path	Gold QA Pairs + Dynamic In-Memory Vector LRU Cache ($N=2048$)	< 0.5 ms	Fallback to full retrieval pipeline on cache miss
🔎 Path 4: Hybrid Search Engine	multilingual-e5-small INT8 ONNX + FAISS HNSW ($M=32$) + RRF ($k=60$) + Script-Aware BM25	< 8.0 ms	Candidate disqualification gate if composite score $< 0.35$
🧠 Path 5: Synthesis & Grounding	Batched IPI Prompt-Guard + Continuous TextRank + SVD Singular Energy + Token Overlap	< 10.0 ms	Return standard non-hallucinating template on grounding fail

---

🌐 Multilingual Indic Language Capability & Provisioning Matrix

VECTOR employs dynamic runtime configuration via config.LANGUAGES as the single source of truth for language federation. The engine provides zero-code hot-swappable expansion across 14 Indic languages and English (~743,000 deduplicated passage records).

ISO Code	Language Target	Script Family	STT Engine Endpoint	Runtime Provisioning Status	Corpus Benchmark Source	Deduplicated Passages
en	English	Latin (Latn)	en-IN	⚡ Active In-Memory Index	MS MARCO English Native Stream	49,507
hi	Hindi	Devanagari (Deva)	hi-IN	⚡ Active In-Memory Index	MS MARCO-XI (hin) Parquet Stream	49,509
mr	Marathi	Devanagari (Deva)	mr-IN	⚡ Active In-Memory Index	MS MARCO-XI (mar) Parquet Stream	49,529
as	Assamese	Bengali/Assamese (Beng)	as-IN	🔌 Zero-Code Hot-Swappable	MS MARCO-XI (asm) Parquet Stream	49,550
bn	Bengali	Bengali (Beng)	bn-IN	🔌 Zero-Code Hot-Swappable	MS MARCO-XI (ben) Parquet Stream	49,531
gu	Gujarati	Gujarati (Gujr)	gu-IN	🔌 Zero-Code Hot-Swappable	MS MARCO-XI (guj) Parquet Stream	49,550
kn	Kannada	Kannada (Knda)	kn-IN	🔌 Zero-Code Hot-Swappable	MS MARCO-XI (kan) Parquet Stream	49,545
ml	Malayalam	Malayalam (Mlym)	ml-IN	🔌 Zero-Code Hot-Swappable	MS MARCO-XI (mal) Parquet Stream	49,542
ne	Nepali	Devanagari (Deva)	ne-NP	🔌 Zero-Code Hot-Swappable	MS MARCO-XI (nep) Parquet Stream	49,520
or	Odia	Odia (Orya)	od-IN	🔌 Zero-Code Hot-Swappable	MS MARCO-XI (ori) Parquet Stream	49,560
pa	Punjabi	Gurmukhi (Guru)	pa-IN	🔌 Zero-Code Hot-Swappable	MS MARCO-XI (pan) Parquet Stream	49,534
sa	Sanskrit	Devanagari (Deva)	sa-IN	🔌 Zero-Code Hot-Swappable	MS MARCO-XI (san) Parquet Stream	49,633
ta	Tamil	Tamil (Taml)	ta-IN	🔌 Zero-Code Hot-Swappable	MS MARCO-XI (tam) Parquet Stream	49,581
te	Telugu	Telugu (Telu)	te-IN	🔌 Zero-Code Hot-Swappable	MS MARCO-XI (tel) Parquet Stream	49,604
ur	Urdu	Perso-Arabic (Arab)	ur-IN	🔌 Zero-Code Hot-Swappable	MS MARCO-XI (urd) Parquet Stream	49,576

Active Memory Allocation: 148,545 Native Passage Vectors + 309 LongDoc Chunks = 148,854 Active Vectors in FAISS HNSW graph memory. Total Available Federated Corpus: ~743,000 Deduplicated Passages.

---

⚡ Enterprise Performance & Benchmark Dashboard

Hardware Environment Specs: 8 vCPUs | 15.78 GB RAM | Windows 11 (AMD64) | 100% CPU Execution
All benchmarks are executed locally on CPU with zero GPU requirement.

Metric	Measured Value	Target SLA Budget	Margin / Performance
🏎️ Total Retrieval Latency	7.04 ms	50.00 ms	⚡ 84.0% Faster than Budget
❄️ Cold-Start 15-Lang Pass Rate	100.0%	< 200.00 ms	✅ 15/15 Languages Passed
🚀 System Throughput	51.7 QPS	—	⚡ 750 Queries in 14.5s
🛡️ Neural Threat Interception	0.24 ms	< 20.00 ms	⚡ Sub-Millisecond Guard

---

1. 🏎️ End-to-End Retrieval Latency Budget SLA (python -m app.benchmark 50)

Measures combined query embedding vectorization (intfloat/multilingual-e5-small INT8 ONNX) + FAISS HNSW graph traversal ($148,545\text{ vectors}$) against the 50ms budget:

STAGE                   P50 LATENCY    PERCENTILE DISTRIBUTION & LATENCY SLAS
─────────────────────────────────────────────────────────────────────────────────────
Query Vectorization     █ 6.21 ms      [ P50: 6.21ms | P95: 7.28ms | P99: 7.94ms ] (INT8 ONNX)
FAISS HNSW Traversal    █ 0.71 ms      [ P50: 0.71ms | P95: 0.93ms | P99: 1.16ms ] (Sub-1ms)
─────────────────────────────────────────────────────────────────────────────────────
TOTAL RETRIEVAL SLA     █ 6.96 ms      [ P95: 7.97ms | P99: 8.95ms | SLA: 50.00ms ] ✅ PASS

Pipeline Retrieval Stage	Avg Latency	P50 (Median)	P95 Latency	P99 Latency	Budget SLA	Status
Query Embedding (multilingual-e5-small ONNX)	6.31 ms	6.21 ms	7.28 ms	7.94 ms	—	⚡ ONNX Accelerated
FAISS HNSW Search (148,545 vectors)	0.73 ms	0.71 ms	0.93 ms	1.16 ms	—	⚡ Sub-1ms Graph Traversal
Total Retrieval Latency (Embed + Search)	7.04 ms	6.96 ms	7.97 ms	8.95 ms	50.00 ms	✅ PASS (84% Faster)

---

2. ❄️ Cold-Start Multilingual SLA Matrix (15 Languages, bypass_cache=True)

Evaluates cold-path retrieval, reranking, context safety scanning, and grounded generation across all 15 languages with cache bypass to guarantee strict SLA compliance:

Language Family	Target Language	Code	Context Guard	Cross-Encoder Rerank	Generation	Total Cold Latency	SLA Status
Indo-Aryan	English	en	0.99 ms	53.40 ms	0.92 ms	120.48 ms	✅ SLA MET
Indo-Aryan	Hindi	hi	1.57 ms	88.12 ms	0.25 ms	175.05 ms	✅ SLA MET
Indo-Aryan	Marathi	mr	1.59 ms	103.31 ms	0.23 ms	177.35 ms	✅ SLA MET
Indo-Aryan	Gujarati	gu	1.82 ms	89.76 ms	0.25 ms	161.61 ms	✅ SLA MET
Indo-Aryan	Punjabi	pa	2.21 ms	100.67 ms	0.19 ms	181.81 ms	✅ SLA MET
Indo-Aryan	Assamese	as	1.83 ms	110.64 ms	0.23 ms	194.22 ms	✅ SLA MET
Indo-Aryan	Odia	or	1.99 ms	86.99 ms	0.23 ms	178.73 ms	✅ SLA MET
Indo-Aryan	Nepali	ne	1.31 ms	109.84 ms	0.23 ms	184.45 ms	✅ SLA MET
Indo-Aryan	Sanskrit	sa	0.00 ms	66.14 ms	0.00 ms	145.40 ms	✅ SLA MET
Dravidian	Tamil	ta	1.35 ms	79.50 ms	0.19 ms	173.49 ms	✅ SLA MET
Dravidian	Telugu	te	1.21 ms	90.72 ms	0.18 ms	164.77 ms	✅ SLA MET
Dravidian	Kannada	kn	1.82 ms	76.86 ms	0.27 ms	167.62 ms	✅ SLA MET
Dravidian	Malayalam	ml	1.82 ms	83.41 ms	0.16 ms	170.34 ms	✅ SLA MET
Perso-Arabic	Urdu	ur	1.58 ms	103.12 ms	0.21 ms	169.99 ms	✅ SLA MET
Bengali-Assamese	Bengali	bn	1.96 ms	93.57 ms	0.17 ms	162.78 ms	✅ SLA MET
System Control	Out-of-Domain	en	0.00 ms	97.39 ms	0.00 ms	168.86 ms	✅ PASS (Declined)
System Control	Prompt Injection	en	0.00 ms	0.00 ms	0.00 ms	0.24 ms	✅ PASS (Blocked)

---

3. 🚀 High-Throughput Speed Benchmark (750 Queries Total)

Throughput: 51.7 Queries / second across 15 Indic languages ($14.50\text{ seconds}$ total execution time):

Pipeline Stage / Metric	P50 (Median)	P70	P90	P99	Mean Latency	Hardware Optimization Mechanism
Query Vectorization	15.18 ms	17.01 ms	22.14 ms	46.44 ms	16.82 ms	ONNX Dynamic Shapes INT8 Quantization
FAISS Graph Search	< 0.90 ms	< 0.90 ms	< 0.90 ms	0.91 ms	0.86 ms	In-Memory HNSW Graph + search_k Slicing
Cross-Encoder Rerank	26.70 ms	108.49 ms	147.18 ms	203.29 ms	108.50 ms	ONNX MiniLM + 64-Token Bounding
Context Synthesis	8.50 ms	8.80 ms	9.20 ms	12.40 ms	8.80 ms	Continuous TextRank + SVD Singular Energy
Cache Fast-Path	0.23 ms	0.28 ms	0.35 ms	0.70 ms	0.35 ms	Dynamic In-Memory Vector LRU Cache
Full Pipeline Latency	16.45 ms	18.27 ms	23.78 ms	57.71 ms	19.22 ms	⚡ Sub-20ms Median Full-Pipeline Execution

---

🌟 Technical Deep-Dive & Engineering Rationales

1. 🛡️ Cascaded 4-Tier Pre-Retrieval Safety Guardrails

• Tier-1: Stem + Flexible-Gap Regex (<0.1 ms): Replaces rigid phrase-literal matching with verb/object root stems and variable word-gap matching (max_gap=4). Handles gerunds ("stealing", "fabricating"), irregular past conjugations ("stole", "hid"), and unlisted adjectives across 15 languages.
• Tier-2: Meta Prompt-Guard 86M Neural Safety (~1.5 ms): ONNX-accelerated Direct Prompt Injection (DPI) and Jailbreak classifier. Includes Unicode confusable unrolling and Base64 decoders. Unhandled exceptions strictly fail safe (is_safe=False, model_failed=True).
• Tier-3: Pre-Retrieval Intent Filter: 6-class intent taxonomy filtering creative writing, suggestion requests, personal advice, planning tasks, roleplay chat, and naming prompt categories before vector search.
• Tier-4: Own-Language Centroid Gate: Computes cosine distance from query embeddings to corpus centroids, requiring own_lang_dist <= threshold * 1.5 to prevent cross-language cluster false positives.

2. ⚡ Script-Aware BM25 + FAISS Vector Search

• In-Memory FAISS HNSW: Built with $M=32$, $efConstruction=200$, $efSearch=64$, delivering 0.73ms CPU search over 148,545 passage vectors.
• Script-Aware Score Fusion: Monolingual queries combine BM25 + dense cosine similarity (HYBRID_BM25_WEIGHT = 0.35). Cross-script queries (e.g. English -> Hindi) automatically detect script mismatch and bypass BM25 lexical penalties.
• Disqualification Gate: Rejects candidate matches under score threshold 0.35 with standard non-hallucinating template.

3. 🧠 Deterministic TextRank + SVD Context Synthesis

• Continuous TextRank Graph Centrality: Computes sentence adjacency matrix $W_{ij} = \max(0, \vec{s}_i \cdot \vec{s}_j)$ with query relevance prior power iterations.
• SVD Matrix Energy Filtering: Retains principal components reaching $\ge 95%$ cumulative singular energy to extract salient facts in $<10\text{ ms}$ on CPU with zero LLM API cost.
• Swappable LLM / SLM Adapter: Optional fallback to Groq / Cerebras APIs (llama-3.3-70b-versatile) or local Qwen SLMs.

---

🚀 Quickstart & Comprehensive Local Setup

⚙️ Prerequisites & System Requirements

• Python: Python 3.10+ (Tested on 3.11 and 3.13)
• System Audio Normalizer: ffmpeg (Required for 16kHz audio preprocessing in STT pipeline)
• RAM Allocation: Minimum 8 GB (16 GB recommended for loading full in-memory 15-language FAISS index)
• Hardware Acceleration: 100% CPU Execution via ONNX Runtime & FAISS-CPU (Zero GPU required)

---

1. 📦 Installation & Environment Setup

# Clone the repository
git clone https://github.com/Rishikvelagapudi/VECTOR.git
cd VECTOR

# Create and activate virtual environment
python -m venv venv
# On Windows PowerShell:
.\venv\Scripts\Activate.ps1
# On Linux / macOS:
source venv/bin/activate

# Upgrade pip and install all Python dependencies
pip install --upgrade pip
pip install -r requirements.txt

---

2. 🔑 Environment Configuration (.env)

Copy the template file .env.example to .env:

cp .env.example .env

Configure your secrets in .env:

# Sarvam AI STT API Key (Saaras v3 Speech Recognition)
SARVAM_API_KEY=your_sarvam_api_key_here

# Primary Generative Provider (Gemini Flash / OpenAI-compatible endpoint)
GEMINI_API_KEY=your_gemini_api_key_here
LLM_API_KEY=your_gemini_api_key_here
LLM_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai/
LLM_MODEL=gemini-2.5-flash

# Hard Safety Overrides & Offline Flags
ALLOW_NETWORK_CALLS_IN_PIPELINE=true
ENABLE_PROMPT_GUARD=true
ENABLE_QUERY_INTENT_FILTER=true

# Embedding & Search Engine Configuration
EMBEDDING_MODEL_NAME=intfloat/multilingual-e5-small

# Local Server Settings
HOST=0.0.0.0
PORT=7860

100% Offline Mode: Setting ALLOW_NETWORK_CALLS_IN_PIPELINE=false forces VECTOR to bypass external LLM API calls and run purely on local CPU ONNX models and deterministic TextRank + SVD energy context synthesis.

---

3. 🖥️ Running Application Interfaces

VECTOR provides 3 distinct execution entrypoints tailored for developers, API integrators, and terminal power users:

Option A: Web Command Center UI (python app.py)

Launches the full retro-tropical Web Audio Command Center UI with live Web Audio frequency canvas and real-time 9-stage telemetry waterfall:

python app.py

Open http://localhost:7860 in your browser.

Option B: High-Speed FastAPI REST Server (uvicorn)

Runs the production REST API exposing /query, /health, and /languages endpoints:

uvicorn api.main:app --host 0.0.0.0 --port 7860 --reload

• Interactive OpenAPI / Swagger Docs: http://localhost:7860/docs
• Health Check Status: GET http://localhost:7860/health
• Active Languages Registry: GET http://localhost:7860/languages

Option C: Interactive Terminal CLI (demo/cli_demo.py)

Run instant terminal queries in text, audio, or interactive shell mode:

# 1. Direct text query with language hint:
python demo/cli_demo.py --text "हृदय के चार कक्ष कौन से हैं?" --lang hi

# 2. Audio file query:
python demo/cli_demo.py --audio sample.wav --lang ta

# 3. Interactive Shell Mode:
python demo/cli_demo.py --interactive

---

4. 🏎️ Running Benchmarks & Verification Suite

# 1. Run full 50-test unit & integration test suite (50/50 passing):
pytest tests/ -v

# 2. Run 50ms retrieval latency budget check (ONNX Embed + FAISS Traversal):
python -m app.benchmark 50

# 3. Run cold-start 15-language SLA benchmark (Cache-bypassed):
python benchmark/run_cold_start_bench.py

# 4. Run 750-query high-throughput speed benchmark (51.7 QPS):
python benchmark/run_speed_bench_50.py

---

5. 🛠️ Data Pipeline & Sample Index Builders

Re-build FAISS HNSW indexes or extract MS MARCO multilingual corpora from scratch:

# Build sample FAISS HNSW indexes locally
python build_sample_indices.py

# Extract and deduplicate MS MARCO corpora for active languages
python data/build_corpus.py

# Extract and build all 15 language corpora streams
python data/build_all_15_corpora.py

---

6. 🐳 Docker Containerization & HF Spaces Deployment

Build and run VECTOR inside a self-contained Docker container:

# Build Docker image
docker build -t vector-rag .

# Run container locally on port 7860
docker run -p 7860:7860 --env-file .env vector-rag

---

🧪 Test Suite & Verification (50/50 Passing)

Run the full automated test suite:

pytest tests/ -v

Test File	Count	Coverage & Scope
tests/test_eval_fixes.py	17	Adversarial safety, intent classification, Prompt-Guard fail-safe, centroid weighting
tests/test_pipeline.py	27	Passage/window/semantic chunking, BM25 script fusion, grounding overlap, 15-lang routing
tests/test_prompt_guard.py	6	DPI injection, IPI context filtering, confusable unpacker, sub-20ms latency check

---

📁 Repository Structure

VECTOR/
├── api/                  # FastAPI web server (/query, /health, /languages)
├── app/                  # Fast ONNX retriever & 50ms benchmark runner
├── benchmark/            # Latency, cold-start, & throughput evaluation scripts
│   └── results/          # JSON & Markdown benchmark reports
├── chunking/             # Native, sentence-window, semantic, & RRF splitters
├── data/                 # FAISS HNSW indexes, centroids, JSONL corpora scripts
├── demo/                 # VECTOR Web Audio UI & visual assets
├── generation/           # TextRank + SVD non-LLM synthesis & LLM fallback
├── guardrails/           # 4-tier cascaded pre-retrieval & post-gen grounding
├── pipeline/             # Async 9-stage pipeline state machine & Pydantic schemas
├── retrieval/            # multilingual-e5-small INT8 ONNX & FAISS engine
├── stt/                  # Sarvam Saaras STT & ffmpeg 16kHz audio pipeline
├── tests/                # 50/50 unit & integration test suite
├── training/             # SFT dataset generator & Qwen Colab training notebooks
├── app.py                # VECTOR Space entrypoint application
├── config.py             # Single source of truth configuration
├── Dockerfile            # Container definition for Hugging Face Spaces
└── requirements.txt      # Python dependencies

---

🚀 Deployment & Live Endpoints

• Live Deployed Application: https://hungry-games-dance.loca.lt
• Hugging Face Space: https://huggingface.co/spaces/rishik1111/visionquest

Free cpu-basic Spaces sleep after 48h inactivity. Initial container spin-up takes 30–90s. Warm runtime operates at ~7–16 ms.

---

📜 License

MIT License. VECTOR Multilingual Indic RAG Engine.