Prism Live RAG

Event-driven streaming retrieval-augmented generation for Samsung Theme 04. Timestamped text chunks simulate evolving ASR hypotheses: the controller can retrieve before the user finishes speaking, while the final answer is grounded in retrieved Cloud/Govt passages.

Samsung hackathon submission

Theme 04 · Binary Bits · SRM Institute of Science and Technology. The final submission tag is PRISM_GENAI_HACKATHON_Y2026.

• Presentation PDF
• Demo video on Google Drive — 4 minutes 52 seconds; publicly readable.
• AI-use disclosure PDF

The presentation and disclosure are included in the tagged commit. The demo video is hosted on Drive and referenced here at the owner's request.

Run the judge dashboard

Install/start Docker, clone this repository, and provide your own Cerebras key:

git clone https://github.com/Aakarsh-Kumar/prism-live-rag.git
cd prism-live-rag
cp .env.example .env
# Edit .env and set CEREBRAS_API_KEY.
docker compose up

On Windows, copy .env.example using your editor/file manager. Open http://localhost:8080 when startup logs show the dashboard URL. Docker automatically pulls the prebuilt image aakarshkumar25/prism-live-rag:judge from Docker Hub, downloads the pinned corpus/index/model release asset, verifies archive and file checksums, and caches it in a persistent volume. No manual asset download, extraction, GPU, host Python, or local build is required. First run pulls the prebuilt image and downloads assets in ~2-4 minutes; subsequent starts are instantaneous. (To build from source instead, append --build). Allocate at least 6 GB RAM to Docker. See judge quickstart.

The pinned judge asset release is public. On 30 September 2026, anonymous download and a fresh-volume Docker bootstrap passed archive/manifest checks and corpus/query validation. Judges do not need GitHub authentication or a local asset archive.

The dashboard provides all 200 accepted test queries, filters and search, streamed partials/revisions, controller reasons, actual search dispatches, returned passages and ranks, citations, provider-call events, latency/cost fields and raw traces. Completed answers can be followed by a presentation request or a delta refinement. “Multi-intent” filtering is a rule preview, not a human gold label.

Design

1. Rule-based streaming controller: Wait / Retrieve / No-Retrieval on each chunk; semantic controllers remain opt-in.
2. Bounded multi-intent decomposition with rule and optional provider paths.
3. BGE-small dense retrieval in LanceDB plus BM25, weighted/nested RRF fusion, and optional cross-encoder reranking. A generative LLM is not the reranker.
4. Evidence-based synthesis, allowlisted citation IDs, abstention and versioned session refinement.

See the architecture brief, pipeline architecture and telemetry schema.

Verification and honest limits

The release verification recorded 275 tests passed, two skipped. Subsequent bootstrap, security, provider and packaging checks passed 59 focused tests on the host and 59 inside Docker; the full suite was not rerun after these changes. The new image built without a local corpus/model build context, installed the pinned archive into a fresh asset volume and completed two cited queries with valid traces. Cached restart succeeded without networking or an archive mount; configured credentials selected provider mode automatically without making an inference request during that startup check.

The completed CPU dashboard audit contains 200 unique executions and 200 valid traces, zero invalid citation IDs, and actual pre-final search in 118/130 eligible cases (90.77%). The final image also completed two queries with Docker networking disabled and no host asset mounts, keys or GPU.

These are execution/timing results, not an overall answer-quality pass:

• Default offline extraction was uncertain or uncited on 59/100 answerable cases. Three of 40 unanswerable cases returned citations. Quality remains open.
• Historical provider faithfulness scores exclude abstentions and refer to earlier runtime snapshots. They do not establish that the latest release clears every gate.
• Input is simulated ASR text, not live audio. Answers arrive whole, not token-streamed.
• Session refinement has controlled tests and targeted corpus checks, not broad semantic accuracy validation.
• Linux AMD64 was exercised; ARM64 and arbitrary judge hardware were not verified.

The benchmark report separates each protocol, runtime version, denominator and failure. Raw generated evidence and checksums ship in submission bundles rather than Git history.

Cerebras routing

Copy .env.example to .env and set CEREBRAS_API_KEY locally. The dashboard uses CEREBRAS_MODEL=gpt-oss-120b, paced at five requests/minute. Routing determines whether a query needs a provider call; not every query does. Keys stay server-side. With a key, the dashboard/API default to automatic provider routing. No-key operation is explicitly provider-free and not the recommended quality demonstration. Provider mode is quota-dependent and slower. Groq/DeepSeek remain available for legacy CLI paths.

Local development

Use Python 3.11–3.13, not host Python 3.14.

python3.13 -m venv .venv
.venv/bin/python -m pip install -e ".[dev]"
.venv/bin/prism-rag validate-data
.venv/bin/prism-rag serve --port 8080

Local execution requires the corpus described in corpus README. Neural retrieval additionally requires model files and a matching index. bash scripts/setup_embeddings.sh selects ONNX Runtime for the available GPU/CPU. prism-rag index builds an index; it is not a startup step for the Docker judge path. Do not replace the full submission index with a small smoke-test index.

.venv/bin/python -m pytest
.venv/bin/prism-rag run-demo --domain cloud
.venv/bin/prism-rag eval --domain cloud --mode retrieval --max-tasks 50
.venv/bin/prism-rag eval --domain cloud --mode streaming --max-tasks 200

Real provider evaluations are explicit and separate from unit tests/CI.

Security and data handling

This is a local demonstration server, not a public production service. Compose publishes only on loopback. Do not expose it through a public tunnel: there is no authentication, and traces contain query text, context and evidence. The server rejects non-loopback Host names and cross-origin run requests, validates JSON bodies, and sends browser security headers. These controls are not a complete security audit. See security policy.

.env, model/index caches, generated evaluation reports, backup corpora and submission archives are ignored by Git. Retain upstream dataset/model licenses when redistributing assets; the raw corpus remains 122,049 passages and the pre-existing cleaned index contains 101,763 rows.

AI-use disclosure

Developed with AI coding assistance under human direction. The human set requirements, approved implementation and verification, and owns final technical decisions. AI assistance contributed code, diagnostics, documentation and executed verification. Provider requests are confined to explicit runtime/evaluation paths.

Repository layout

• src/prism_live_rag/: pipeline, providers, telemetry and dashboard assets.
• tests/: deterministic unit and regression checks.
• scripts/: preparation, audits, packaging and launch helpers.
• docs/: design references, dated research notes and versioned evaluation reports.
• data/: tracked query/stream fixtures and local ignored corpus/evaluation files.