Vernacular RAG Voicebot for Government Schemes

The prototype treats scheme discovery as a question-answering problem: a person can describe a need or eligibility situation without first knowing the scheme name or source document. English, Hindi, and Bengali text access keeps that search from depending on an English-only query.

Hindi and Bengali voice input serves the same path when typing is not the preferred mode, and vernacular speech is returned only for voice turns. Language choice is explicit because the IndicConformer speech service requires the caller to declare the spoken language rather than relying on automatic detection.

After normalization, multilingual-e5 semantic search covers paraphrase and intent while BM25 protects exact scheme names, eligibility terms, and distinctive wording. Reciprocal-rank fusion combines their ranked candidates without pretending their raw scores share one scale; bge-reranker-v2-m3 then evaluates the fused query-passage pairs more closely.

The reranked set passes only when its strongest score reaches 0.55 and leads the second result by at least 0.03. Weak evidence routes to a fallback; accepted evidence enters a grounded prompt and then a verifier that also fails closed. General conversation deliberately bypasses retrieval and rejoins the path at vernacular translation.

LangGraph checkpoints retain conversation content and graph state under the session thread ID. A separate Alembic-managed chat_sessions table keeps the title, timestamps, and turn count needed to list conversations. The browser retains the active ID, loads the session list, and asks the replay endpoint for the selected history, so resume behavior does not make product metadata part of graph internals.

Evaluation stays outside the request path because the local judge and metric calls are a repeatable regression harness, not user-facing work. Retrieval precision and recall use all 100 runs per language because retrieval happens before gating; answer-quality metrics use only the 84 English, 79 Hindi, and 74 Bengali answered rows; answered rate uses all runs so refusal coverage remains visible.

The fixed three-language replay shows that the final configured pipeline retrieves useful context and keeps answered responses grounded across English, Hindi, and Bengali within this synthetic corpus-derived testset. It also makes the coverage difference visible instead of hiding refusals inside answer-quality averages. It does not establish production usage, open-domain generalization, or quality outside these 300 language-runs.

The most informative next experiment is a human-reviewed parallel Hindi and Bengali set written natively rather than machine-translated, scored alongside the current English-normalized path. That comparison would separate translation effects from retrieval behavior and test whether the same confidence thresholds preserve the quality-versus-coverage balance for native vernacular questions.