Challenge · 01
Local inference
CPU constraints dominate generation latency.
<- AI Systems Dossier
ENTRY NO. 06Research Intelligence / R&DLocal R&DFAISS / MiniLM / Qwen / FastAPILOCAL RAG / R&D
AI Research Intelligence
Identity · CPU-first research assistant
Local RAG for AI research. Search papers, rerank evidence, synthesize source-backed answers.
FAISS / MiniLM / Qwen / FastAPI
Evidence boundary
What is demonstrated, and where the claim stops.
Repository factConfiguration
What this proves
- /A coherent CPU-first retrieval and synthesis architecture is implemented.
- /Exact FAISS retrieval and cross-encoder reranking have inspectable code paths.
- /The system can run without a paid inference API when external artifacts are installed.
What this does not prove
- /Published retrieval or answer quality.
- /A reproducible latency result across hardware.
- /A self-contained checkout with model, corpus, and index included.
Data boundary
Model, corpus, and index are external downloads. Checked-in data directories do not include a runnable corpus or index artifact.
Evidence type
repository fact / configuration
Model, corpus, and index are external downloads. Quality evaluation is currently heuristic and no benchmark result is published.
Known limitations
- /External model and data downloads are required.
- /Current answer evaluation is heuristic.
- /Semantic quality metrics and saved benchmark results are not implemented.
What production would require
- /Governed and updateable research ingestion.
- /A labelled evaluation set with saved retrieval and answer-quality results.
- /Artifact/version management, monitoring, and an operated serving environment.
Problem
A local research assistant needs reproducible retrieval, reranking, and source-bearing synthesis on commodity hardware.
System Flow
5 stages · click to inspect
->
->
->
->
Click any stage to inspect
Signature Module · CPU-first research assistant
Inspect local retrieval
Follow arXiv metadata through topic mapping, MiniLM embeddings, exact FAISS retrieval, reranking, and local synthesis.
Research Intelligence · local R&Dinspectable
INGESTarXiv metadata · deterministic topics
EMBEDMiniLM · 384-dimensional configuration
SEARCHFAISS exact retrieval
RERANKMS MARCO cross-encoder
OUTPUTlocal Qwen synthesis + sources
BOUNDexternal artifacts · no published benchmark
Demonstrated Change
Demo scopeBefore
Research search and synthesis depend on disconnected tools or hosted APIs.
->
Demonstrated build
One local-first R&D pipeline with exact retrieval, reranking, and source-bearing output.
The R&D repository demonstrates the architecture and interfaces for that local-first workflow.
Model, corpus, and index are external downloads. Quality evaluation is currently heuristic and no benchmark result is published.
Proof Signals
Classified evidence12
deterministic research topics
Repository fact
25
defined evaluation cases; results not yet published
Repository fact
0
paid inference APIs required at runtime
Configuration
Every signal is classified as a repository fact, recorded test, configuration, or disclosed data boundary.
Why This Is Hard
3 engineering challengesChallenge · 02
Artifact setup
model, corpus, and index must remain version-aligned.
Challenge · 03
Evaluation
a defined test set is not a quality result without an executable saved run.
Engineering Depth
3 topics · click to expandThe active path uses normalized MiniLM embeddings and FAISS IndexFlatIP before cross-encoder reranking.
Key numbers
384-dimMiniLM embedding configuration
2query modes: fast and full
Built withPythonFAISSMiniLMCross-encoderQwenFastAPIStreamlit