Anantha answers ocean-data questions by combining natural-language routing, semantic retrieval and PostgreSQL queries.
Oceanographic records often require both conceptual discovery and exact filtering. Anantha uses vector retrieval to identify relevant float records when a question is semantic, then retrieves structured observations through SQL. The response layer can produce text, tables and plot data from those results.
Preview · Quick start · Architecture · Development · Contributing
Archived client view showing session navigation, response modes and the ocean-data query composer.
| Workflow | Result |
|---|---|
| Interpret a question | Classify the request into structured or semantic retrieval. |
| Retrieve observations | Use Chroma candidate identifiers to constrain PostgreSQL access. |
| Present the result | Return a narrative, table or plot-oriented response from the retrieved data. |
Provision PostgreSQL, source observations, a populated Chroma collection and provider credentials. The vector module currently uses an absolute local Chroma path and reads GEMINI_API_KEY2; replace the path for your environment before launching.
From a fresh checkout:
git clone https://github.com/ryash14/Anantha.git
cd Anantha
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r backend/requirements.txt
cd backend
python -m uvicorn main:app --host 127.0.0.1 --port 8000
# In another terminal at the repository root: npm install && npm run devConfigure PostgreSQL, source data, Chroma indexes and provider credentials before querying. Inspect the backend database and embedding modules for expected local paths and configuration. The code routes through Python functions; LangGraph and LSTM deployment are not asserted here. Generated SQL requires a restricted database role and validation before exposure to untrusted users.
Before testing /query, inspect these integration points:
| Component | Setup responsibility |
|---|---|
backend/store_in_vector_db/vector_db.py |
Replace the machine-specific Chroma path and populate the documents collection |
GEMINI_API_KEY2 |
Configure the key read by the embedding module |
| PostgreSQL retrieval module | Match connection settings and schema to the observation data |
| Query-generation modules | Configure their provider access and validate generated SQL |
| Browser API helper | Point the client at this backend and allow its origin |
Use a database role restricted to the intended reads. A successful LLM response is not sufficient evidence that a generated SQL statement is safe or that its oceanographic interpretation is correct.
- Submit a question and response mode to the FastAPI
/queryendpoint. - Query enhancement and classification choose a direct SQL path or a vector-assisted path.
- Semantic retrieval applies metadata filters and supplies float IDs to SQL generation.
- PostgreSQL results are converted into the requested answer or data presentation.
flowchart TB
Query["Question and output mode"] --> Route["Enhance and classify"]
Route -->|structured| SQL["SQL generation"]
Route -->|semantic| Vector["Metadata filters and Gemini embedding"]
Vector --> Chroma[("Chroma: candidate float IDs")]
Chroma --> SQL
SQL --> PG[("PostgreSQL observations")]
PG --> Answer["Synthesize text, table or plot data"]
Answer --> API["FastAPI response"]
The orchestrator is ordinary Python routing rather than a LangGraph state machine. Chroma holds semantic document records; PostgreSQL holds structured measurements. Gemini provides embedding and generation calls in the inspected modules.
The repository reflects a research/hackathon implementation: local data paths, database setup and model-provider configuration require adaptation. Forecasting or LSTM production behavior is not established by the request path documented here.
Stack: Python · FastAPI · ChromaDB · PostgreSQL · provider LLM calls · pandas · React · Vite.
- backend/main.py
- backend/main_pipeline.py
- backend/store_in_vector_db/vector_db.py
- backend/retrieve_data_from_db/postgres_db.py
There is no complete automated verification command documented for this application. Start the affected component, use a small synthetic input and inspect its output before changing integrations. The contribution guide describes the evidence expected in a pull request.
Report reproducible bugs or propose changes through GitHub Issues. Include the relevant environment and a small example rather than private datasets or credentials. See contributing, security reporting and the code of conduct.
No repository-wide license is currently provided. Third-party components, models and datasets retain their own terms.