Spaces:
Running
title: Synapse Agentic AI Assistant
emoji: 🧠
colorFrom: indigo
colorTo: green
sdk: docker
app_port: 7860
pinned: true
license: mit
short_description: Agentic AI assistant with autonomous tool use and hybrid RAG
tags:
- agents
- rag
- tool-calling
- fastapi
- react
- gemini
Synapse
An agentic AI assistant with autonomous tool use, hybrid retrieval, and full cost observability.
Click Launch demo in the app above. No signup, no API key. You get a private sandbox preloaded with a sample report so document retrieval works from your first message.
Try these in the demo
| Prompt | What it demonstrates |
|---|---|
What is the weather in Amsterdam right now? |
The agent decides on its own to call a live weather tool, then formats the result |
What was Northwind's 2026 revenue and gross margin? |
Hybrid RAG over the preloaded report, with the source cited inline |
Summarize the autonomy research section |
Multi-step retrieval: the agent issues several searches before answering |
Search the web for recent work on agentic RAG |
Live web search through the tool layer |
What is (1.07 ** 30) * 25000? |
Precise arithmetic through a sandboxed calculator, not token prediction |
Then open Analytics in the sidebar for per-message token, cost and latency tracking, and Documents to upload a PDF of your own.
What makes it interesting
The agent loop is hand-written. No LangChain, no agent framework. About 150 lines of explicit orchestration: stream the model, detect tool calls, execute them concurrently, feed results back, iterate until the model answers. Owning the loop means exact control over event ordering, token accounting across iterations, and graceful degradation when a tool fails.
Retrieval is hybrid, not just vectors. Dense embedding search (ChromaDB) runs alongside BM25 keyword search. Results are combined with Reciprocal Rank Fusion, then optionally reranked by a small model. Dense-only retrieval misses exact identifiers; BM25-only misses paraphrases. RRF needs no score normalization between the two, which makes it robust.
Streaming is a typed protocol. One SSE channel carries token,
tool_start, tool_end, citations, usage, title, done and error
events, so the UI can show tool activity live rather than freezing until the
answer arrives.
Cost is measured, not guessed. Exact token usage is captured per request, accumulated across every iteration of an agent turn, priced from a rate table, and persisted with latency for the analytics dashboard.
Architecture
flowchart LR
UI["React SPA, SSE stream parser"] --> API["FastAPI"]
API --> LOOP["Agent orchestrator"]
LOOP --> TOOLS["Tools: web search, weather, calculator, datetime, document search"]
LOOP --> LLM["OpenAI-compatible LLM API"]
TOOLS --> RRF["RRF fusion + rerank"]
RRF --> CHROMA[("ChromaDB")]
RRF --> BM25["BM25 index"]
API --> DB[("SQLite, async SQLAlchemy")]
Stack: FastAPI, async SQLAlchemy 2, ChromaDB, rank-bm25, PyJWT, React 18, Vite, Recharts. Model access goes through an OpenAI-compatible provider layer, so OpenAI, Gemini, Groq or a local Ollama server are a two-variable change. This Space runs on Gemini's free tier.
Notes on this deployment
- Guest sessions are isolated from each other and cleared periodically.
- Fair-use limits apply (messages per hour per visitor, plus a daily ceiling) because the demo runs on one shared free-tier key.
- The Space container has ephemeral storage, so the database resets on restart. The sample document is re-seeded automatically on boot.
- For unlimited use, clone the GitHub repo and run it locally with your own free Gemini key.
Built by Adwitiya Shukla. Source, tests and architecture notes: github.com/adwitiyashukla/synapse