Meridian

How it behaves

For questions that need fresh or grounded information, Meridian can emit tool calls like:

xml
Copy code
<call>
{"tool":"browser.search","input":{"query":"latest OpenAI ChatGPT model announcement"}}
</call>

After your app executes the tool and returns source text, Meridian is expected to answer with citations:

xml
Copy code
<response>
OpenAI announced GPT-5.2 as a newer frontier model for ChatGPT and API users, highlighting stronger reasoning, long-context understanding, coding, and vision capabilities. [source:1]
</response>

Your app is responsible for actually running the tools. The model only decides when and how to call them.

Recommended app loop

1. Send the user message and system prompt to Meridian.
2. Watch the generation for <call>...</call> blocks.
3. Parse the JSON inside each call.
4. Execute the matching tool in your app.
5. Feed the tool result back to the model as source text.
6. Ask the model for the final <response>.

Example source block:

xml
Copy code
<tool_result>
{"source_id":1,"title":"Example article","url":"https://example.com","text":"Relevant page text..."}
</tool_result>

Prompt style

A good system prompt is direct and tool-aware:

text
Copy code
You are Meridian, a source-grounded research assistant.
Use browser.search when current or factual information is needed.
Use browser.open to read promising sources.
Answer concisely and cite sources inline as [source:1], [source:2].
If sources do not support a claim, say so.
Do not invent citations.

What Meridian is good at

• research assistant workflows
• live search and browsing agents
• citation-style answers
• tool-call formatting
• short factual summaries from provided sources
• app-integrated RAG experiences

What Meridian is not

• a search engine by itself
• a guarantee of truth without retrieval
• a replacement for source filtering or ranking
• a model that can browse unless your app gives it browsing tools

Safety and reliability notes

Meridian should be used with source validation in your application. For best results:

• prefer trusted domains when ranking search results
• open and read sources before asking for the final answer
• require citations for factual claims
• show source links to users
• handle tool errors explicitly
• avoid letting the model invent source IDs

Loading

This repository contains a merged Transformers model. Load it with transformers:

python
Copy code
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Pharos-ai/Meridian"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

For production use, connect Meridian to your own search, browser, and retrieval tools.