rishik1111 commited on
Commit
5bb2554
Β·
1 Parent(s): 3354ccf

docs: redesign System Architecture to Horizontal Swimlane Architecture with custom styling and execution matrix

Browse files
Files changed (1) hide show
  1. README.md +63 -29
README.md CHANGED
@@ -44,39 +44,73 @@ The active runtime deployment loads **148,854 in-memory FAISS vectors** across 3
44
 
45
  ---
46
 
47
- ## πŸ›οΈ System Architecture
48
 
49
  ```mermaid
50
- graph TD
51
- A[Spoken Voice Audio / Text Bypass] --> B[Sarvam Saaras v3 STT + ffmpeg 16kHz Normalizer]
52
- B --> C[VECTOR Language Resolution Router: config.LANGUAGES]
53
- C --> D[VECTOR Guardrail 1: Tier-1 Fast Regex + Safety Patterns]
54
- D -- Safe --> PG[VECTOR Guardrail 2: Meta Prompt-Guard 86M Neural DPI Shield]
55
- D -- Blocked --> X[Declined Response: Safety Violation]
56
- PG -- Safe --> IF[VECTOR Guardrail 3: Pre-Retrieval Query Intent Filter]
57
- PG -- Injected --> X
58
- IF -- Factual --> E[Query Embedding: 'query: ' Prefix multilingual-e5-small INT8]
59
- IF -- Non-Factual Intent --> X
60
- E --> F[VECTOR Guardrail 4: Centroid Distance Off-Topic Filter]
61
- F -- Off-Topic --> X
62
- F -- On-Topic --> CACHE{Dynamic Vector & Gold QA Cache}
63
- CACHE -- Cache Hit <0.5ms --> N[Grounded Response + Zero-Latency Fast Path]
64
- CACHE -- Cache Miss --> G[VECTOR Parallel Multi-Strategy FAISS HNSW Engine]
65
- G --> H1[Passage Native Index: 148,545 Vectors]
66
- G --> H2[Semantic LongDoc Index: 309 Vectors]
67
- H1 --> I[Candidate Merge & Reciprocal Rank Fusion RRF k=60]
68
- H2 --> I
69
- I --> J[VECTOR Adaptive Script-Aware BM25 Score Fusion]
70
- J --> K[Relevance & Disqualification Gate: Dense / CE Threshold]
71
- K -- Score < Threshold --> Y[Declined Response: No Relevant Info in Corpus]
72
- K -- High Relevance --> CS[Context Chunk Safety: Batched Prompt-Guard 86M IPI Scan]
73
- CS -- Poisoned Chunks --> X
74
- CS -- Clean Chunks --> L[VECTOR Deterministic Context Synthesis: TextRank + SVD Energy]
75
- L --> M[VECTOR Post-Generation Grounding & Hallucination Guardrail]
76
- M -- Grounded --> N[Grounded JSON Response + Full 9-Stage Telemetry]
77
- M -- Insufficient Info --> Y
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
78
  ```
79
 
 
 
 
 
 
 
 
 
 
 
80
  ---
81
 
82
  ## 🌐 Language Extensibility Matrix
 
44
 
45
  ---
46
 
47
+ ## πŸ›οΈ Horizontal Swimlane Architecture
48
 
49
  ```mermaid
50
+ graph LR
51
+ %% Custom Styling Classes
52
+ classDef inputStyle fill:#1E293B,stroke:#38BDF8,stroke-width:2px,color:#F8FAFC;
53
+ classDef sttStyle fill:#0F172A,stroke:#818CF8,stroke-width:2px,color:#F8FAFC;
54
+ classDef guardStyle fill:#311B92,stroke:#B388FF,stroke-width:2px,color:#FFFFFF;
55
+ classDef cacheStyle fill:#064E3B,stroke:#34D399,stroke-width:2px,color:#F8FAFC;
56
+ classDef faissStyle fill:#164E63,stroke:#22D3EE,stroke-width:2px,color:#F8FAFC;
57
+ classDef synthStyle fill:#4C1D95,stroke:#C084FC,stroke-width:2px,color:#F8FAFC;
58
+ classDef outStyle fill:#065F46,stroke:#10B981,stroke-width:2px,color:#FFFFFF;
59
+ classDef blockStyle fill:#881337,stroke:#F43F5E,stroke-width:2px,color:#FFFFFF;
60
+
61
+ subgraph PATH1 ["πŸŽ™οΈ Path 1: Audio & Text Ingestion"]
62
+ A[Audio Upload / Microphone Stream] ::: inputStyle --> STT[Sarvam Saaras STT + ffmpeg 16kHz] ::: sttStyle
63
+ T[Raw Text Input Bypass] ::: inputStyle --> ROUTER[Language Resolution Router] ::: sttStyle
64
+ STT --> ROUTER
65
+ end
66
+
67
+ subgraph PATH2 ["πŸ›‘οΈ Path 2: 4-Tier Security Shield"]
68
+ ROUTER --> G1[Tier-1 Stem Regex + Obfuscation Decoder] ::: guardStyle
69
+ G1 -- Safe --> G2[Tier-2 Meta Prompt-Guard 86M DPI] ::: guardStyle
70
+ G2 -- Safe --> G3[Tier-3 6-Class Query Intent Gate] ::: guardStyle
71
+ G3 -- Factual --> G4[Tier-4 Own-Lang Centroid Distance Gate] ::: guardStyle
72
+ end
73
+
74
+ subgraph PATH3 ["⚑ Path 3: Sub-0.5ms Hot Cache Fast-Path"]
75
+ G4 -- On-Topic --> CACHE{Hot Cache Lookup} ::: cacheStyle
76
+ CACHE -- "Hit (<0.5ms)" --> FAST_OUT[Zero-Latency Response] ::: outStyle
77
+ end
78
+
79
+ subgraph PATH4 ["πŸ”Ž Path 4: Hybrid Vector Retrieval Engine"]
80
+ CACHE -- "Miss" --> EMB[multilingual-e5-small INT8 ONNX] ::: faissStyle
81
+ EMB --> FAISS[Parallel FAISS HNSW Native & LongDoc Search] ::: faissStyle
82
+ FAISS --> RRF[Reciprocal Rank Fusion k=60] ::: faissStyle
83
+ RRF --> BM25[Adaptive Script-Aware BM25 Fusion] ::: faissStyle
84
+ BM25 --> GATE{Disqualification Gate} ::: faissStyle
85
+ end
86
+
87
+ subgraph PATH5 ["🧠 Path 5: Deterministic Synthesis & Grounding"]
88
+ GATE -- High Relevance --> IPI[Batched Prompt-Guard IPI Context Scan] ::: synthStyle
89
+ IPI -- Clean Chunks --> SYNTH[Continuous TextRank + SVD Energy Synthesis] ::: synthStyle
90
+ SYNTH --> GROUND[Post-Gen Grounding Overlap Verifier] ::: synthStyle
91
+ GROUND -- Grounded --> FINAL_OUT[JSON Response + 9-Stage Telemetry] ::: outStyle
92
+ end
93
+
94
+ %% Rejection Routing
95
+ G1 -- Blocked --> REJECT[Declined Response: Safety Violation] ::: blockStyle
96
+ G2 -- Injected --> REJECT
97
+ G3 -- Non-Factual --> REJECT
98
+ G4 -- Off-Topic --> REJECT
99
+ GATE -- Score < 0.35 --> DECLINE[Declined Response: Insufficient Info] ::: blockStyle
100
+ IPI -- Poisoned --> REJECT
101
+ GROUND -- Ungrounded --> DECLINE
102
  ```
103
 
104
+ ### πŸ›£οΈ Swimlane Pipeline Execution Breakdown
105
+
106
+ | Swimlane / Path | Key Components & Models | Latency Budget | Action on Failure / Edge Case |
107
+ | :--- | :--- | :---: | :--- |
108
+ | **πŸŽ™οΈ 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 |
109
+ | **πŸ›‘οΈ 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 |
110
+ | **⚑ 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 |
111
+ | **πŸ”Ž 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 $< 0.35$ |
112
+ | **🧠 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 |
113
+
114
  ---
115
 
116
  ## 🌐 Language Extensibility Matrix