praveen-goml/genaitask

Purpose and entry points

JARVIS is a Python assistant backend combining retrieval-augmented generation (RAG), conversational memory, and guarded natural-language-to-SQL. It exposes an HTTP API, plus CLI and Streamlit interfaces.

• API application: JARVIS/src/api/main.py

• Interactive clients: JARVIS/src/ui/cli.py, JARVIS/src/ui/streamlit_app.py

• API route surface: JARVIS/src/api/routes/chat.py, JARVIS/src/api/routes/sql.py, JARVIS/src/api/routes/health.py

API and session boundary

The API layer validates request/response payloads, dispatches chat and SQL requests, and provides a health endpoint. Session state is separated from request routing, allowing chat continuity to be managed independently of the transport layer.

• Schemas: JARVIS/src/api/schemas.py

• Session handling: JARVIS/src/api/session_store.py

• Chat route: JARVIS/src/api/routes/chat.py

Agent orchestration

The core agent is the central control point for user interactions. It coordinates the LLM, RAG context, memory, and available tools, selecting or composing these capabilities for a response rather than embedding that logic in UI or route code.

• Agent: JARVIS/src/core/agent.py

• LLM integration boundary: JARVIS/src/core/llm.py

• Tool integration: JARVIS/src/core/tools.py

Retrieval-augmented knowledge flow

Documents under data/documents are ingested, chunked, embedded/indexed, and retrieved to ground responses. Retrieval is isolated from agent orchestration, while chunking and vector persistence are separate concerns.

• Ingestion entry point: JARVIS/scripts/ingest.py

• RAG pipeline: JARVIS/src/core/rag.py

• Chunking and vector storage: JARVIS/src/core/chunking.py, JARVIS/src/core/vector_store.py

• Source corpus: JARVIS/data/documents

Conversational memory

Conversation memory is a distinct core subsystem, likely used by the agent alongside retrieved documents. This separates transient/session-level chat context from the durable knowledge corpus and relational database.

• Memory component: JARVIS/src/core/memory.py

• Agent consumer: JARVIS/src/core/agent.py

• API session boundary: JARVIS/src/api/session_store.py

Relational data and NL-to-SQL safety pipeline

Natural-language SQL requests follow a dedicated pipeline: inspect available schema, generate SQL, validate/guard it, then execute against the database layer. This is a material safety boundary: SQL generation is not directly coupled to API routing or agent internals.

• Pipeline and generator: JARVIS/src/nl2sql/pipeline.py, JARVIS/src/nl2sql/generator.py

• Guard and schema inspection: JARVIS/src/nl2sql/guard.py, JARVIS/src/nl2sql/schema_introspection.py

• Database models/access: JARVIS/src/db/database.py, JARVIS/src/db/models.py

Runtime, initialization, and infrastructure

The repository is a Python service with dependency definitions and container orchestration. Database setup and seed data are explicit operational stages, separate from document ingestion.

• Dependencies: JARVIS/requirements.txt

• Container runtime composition: JARVIS/docker-compose.yml

• Database lifecycle scripts: JARVIS/scripts/setup_db.py, JARVIS/scripts/seed_db.py

• Environment configuration: JARVIS/config/settings.py, JARVIS/.env.example