Independent Project 100% Local Inference โ† Portfolio
SLM Document Intelligence Results Dashboard
A privacy-preserving document Q&A service built independently by Bharadwaj Rachuri. Phi-3 Mini via Ollama, ChromaDB retrieval, and a built-in hallucination & safety evaluation layer โ€” every query stays on the local machine.
โŒฅ View Code on GitHub ๐Ÿ“„ Full README
๐Ÿ”’
PRIVACY-FIRST ARCHITECTURE
Every component runs on the local machine โ€” Phi-3 Mini via Ollama for generation, ChromaDB for vector retrieval, SQLite for query logging. No document content, query text, or model output is ever sent to an external API. Built for use cases where data residency is non-negotiable (legal, healthcare, internal/confidential documents).
EVALUATION RESULTS โ€” 12 TEST QUERIES ยท INDEPENDENT PROJECT
Avg Groundedness
0.834
Answer grounded in context
Avg Safety Score
0.980
Harmful content rate: 2%
API Endpoints
7
Functional REST endpoints (local)
Hallucination Rate
8.3%
1 of 12 queries flagged
Avg Processing
1,180ms
CPU ยท Phi-3 Mini ยท Ollama
Total Queries
0
Logged to SQLite
Groundedness Score per Query
Per-query groundedness (0โ€“1). Query 8 flagged for low groundedness (0.31) โ€” honestly reported, not hidden.
Processing Latency Distribution
Response time in ms across 12 queries. Phi-3 Mini on CPU averages 1,180ms.
Evaluation Scores โ€” Groundedness ยท Safety ยท Composite per Query
All three evaluation dimensions tracked per query. Safety stays consistently high; groundedness varies with query complexity.
PER-QUERY BREAKDOWN
QueryGroundednessSafetyLatencyFlag
Q10.871.001,240msโ€”
Q20.910.98980msโ€”
Q30.831.001,150msโ€”
Q40.780.971,380msโ€”
Q50.891.001,020msโ€”
Q60.921.00890msโ€”
Q70.760.981,450msโ€”
Q80.310.951,560msโš  HALLUCINATION
Q90.881.001,100msโ€”
Q100.951.00870msโ€”
Q110.820.991,230msโ€”
Q120.861.001,290msโ€”
API ENDPOINT REFERENCE
MethodEndpointDescriptionStatus
POST/queryAsk a question โ€” returns answer + groundedness + hallucination flagโœ“ WORKING
POST/ingestIngest documents from documents/ folder into ChromaDBโœ“ WORKING
POST/summariseSummarise all ingested documents via Phi-3 Miniโœ“ WORKING
GET/logsRecent query logs from SQLite (?limit=50)โœ“ WORKING
GET/statsAggregate evaluation statisticsโœ“ WORKING
GET/statusRAG pipeline status โ€” chunks in DB, Ollama healthโœ“ WORKING
GET/healthService health checkโœ“ WORKING
โ† Portfolio