ChronosGuru
Persona-driven historical RAG platform hitting 92% factual accuracy, holding persona fidelity across 20+ turns, with hybrid keyword + semantic search cutting retrieval latency under 200ms.
Pipeline
The problem
History teaching leans on lecture and memorisation, so learners retain dates rather than the reasoning and stakes behind events — and a plain chatbot playing a historical figure drifts out of character and invents facts.
Architecture
End-to-end flow, stage by stage.
- 01
Document ingest
- 02
Hybrid search
- 03
Vector DB
- 04
Persona prompt
- 05
Murf TTS
How it works
Conversation instead of recitation
Learners talk directly to historical figures, turning passive reading into dialogue where motivations and consequences can be questioned.
Hybrid retrieval for factual grounding
Keyword and semantic search run together over a curated vector store so answers come from verified source material, lifting factual accuracy to 92% while retrieval completes in under 200ms.
Persona fidelity over long sessions
Dynamic system prompts plus stateful session memory hold each figure's voice and worldview past 20 turns, where a static prompt collapses after a handful.
Voice and delivery
Claude drives the dialogue and Murf text-to-speech gives each figure a spoken voice inside a Next.js and Tailwind interface deployed to Vercel.
- Grounded historical conversations in a curated vector database, combining keyword and semantic retrieval to answer from verified source material rather than model memory alone.
- Used dynamic system prompts and stateful session memory to preserve the voice and worldview of each historical figure across long conversations.
- Delivered sub-200 ms retrieval, 92% factual accuracy, and persona consistency beyond 20 turns in a Next.js experience with Claude and Murf voice output.
By the numbers
Grounding and persona gains
Hybrid retrieval and stateful prompting against a plain prompted chatbot.
Factual accuracy (%)
higher is betterRetrieval latency (ms)
lower is betterTurns holding persona
higher is betterWhat it ships with
- Talk to historical figures in character
- Hybrid keyword plus semantic retrieval
- Session memory for long conversations
- Murf voice narration per persona
- Live hosted demo
- 72%→92%
- Factual accuracy
- ~1.2s→<200ms
- Retrieval latency
- 3 turns→20+
- Turns of persona fidelity
Stack
More projects
Agentic AI
Pitch Deck Debater
Simulated investor boardrooms with VC, CTO, and Product Lead personas, an adversarial Chain-of-Thought feedback loop, and consensus-voting logic that resolves inter-agent hallucinations.
Edge / IoT
On-Device Small Language Model Chatbot
Fully offline SLM chatbot running on-device with no cloud dependency — quantized inference tuned for constrained memory and latency budgets.