Source Code Walkthrough

Source Code Walkthrough

Recommended entry point

RAG_Service_Corrected_2026_Diagnostics.py implements the most useful command-line switch set and has an extra run_doctor() dependency-diagnostics routine. Start here; the notebook copies its logic into separate teaching cells.

Read the file top to bottom

Approximate section	Symbol	Explanation
Constants / data	COLLECTION, MODEL_ID, DOCUMENTS	Qdrant name, embedding checkpoint, fictional KB
Contracts	RAGRequest, RAGResponse	Parameter validation and output shape
Privacy	PIIFilter	Regex rules, Luhn validation, IPv4 validation, masking
Input checks	PromptInjectionDetector	Fixed expression matching and risk-score heuristic
Caching	CacheEntry, SemanticCache	Scope, TTL, exact lookup, embedding similarity, eviction
Metrics	SystemMonitor	Synchronized counters, elapsed time, optional RAM
Retrieval core	RAGService	Dependencies, bootstrap collection, query and answer flow
HTTP	create_app(service)	FastAPI, tenant auth, Pydantic body, SlowAPI, routes
Self-test	_run_self_test()	Four unittest checks; no real model download
Diagnostics	run_doctor()	Package versions/import checks/GPU availability
Bootstrap	main()	Parse --doctor, --self-test, --demo, --api

Embedding and collection startup

RAGService.__init__ optionally accepts a test encoder, Qdrant client and documents. Without them it loads SentenceTransformer(MODEL_ID, device=EMBED_DEVICE), selects QdrantClient(":memory:") unless QDRANT_PATH is set, initializes components and calls _initialize_collection().

_initialize_collection() embeds each stored passage, creates the collection if missing with cosine distance, and upserts PointStruct records whose payload includes tenant ID, category and text. The source contains no incremental ingestion endpoint.

Retrieval

retrieve() encodes a cleaned query, calls client.query_points with the tenant filter, similarity threshold and requested result count, and returns dictionaries with id, score and payload.

_extractive_stub() joins passage texts into the response. No prompt is sent to an LLM.

CLI and notebook

• --self-test: regex behavior, false positives for invalid cards, cache scope and request validation.
• --doctor: dependency import/version and CUDA checks (no model download).
• --demo: built-in sample queries; prints cache state, sources and monitor report.
• --api: starts local FastAPI server after embedding model and sample KB initialize.
• Notebook: nearly the same code broken into cells; real model demo opt-in via RUN_REAL_DEMO.

Alternate versions

RAG_Service_Corrected_2026_Windows.py and rag_fixed_v2.py use a similar structure but do not include --doctor. Keep them as historical/alternative demos, not as three simultaneously deployed services.

See Low-Level-Design for algorithms and Testing-and-Validation for reproducible checks.