Please address the comments from this code review:

## Overall Comments
- There’s a lot of duplicated logic across PrepareRequest, PrepareRequestWithSchema, PrepareStreamRequest, and PrepareRequestWithMessages—consider extracting a shared helper for building/marshaling requests and merging options to reduce repetition.
- NewVLLMProvider currently just concatenates baseURL without validation—add URL normalization (ensure scheme, trailing slash or v1 path) to avoid malformed endpoints at runtime.
- HandleFunctionCalls is unimplemented but consumers might expect it—either implement function-calling support or explicitly document/disable it in the provider capabilities to prevent unexpected errors.

## Individual Comments

### Comment 1
<location> `providers/vllm.go:99` </location>
<code_context>
+return headers
+}
+
+// PrepareRequest creates the request body for a vLLM API call.
+func (p *VLLMProvider) PrepareRequest(prompt string, options map[string]interface{}) ([]byte, error) {
+request := map[string]interface{}{
</code_context>

<issue_to_address>
**issue (complexity):** Consider refactoring request-building logic into a helper method to eliminate code duplication and simplify maintenance.

```go
// Add this helper in VLLMProvider to consolidate request‐building/option merging:
func (p *VLLMProvider) buildRequest(
    messages []map[string]interface{},
    opts map[string]interface{},
    extras ...func(map[string]interface{}),
) ([]byte, error) {
    req := map[string]interface{}{
        "model":    p.model,
        "messages": messages,
    }
    // merge provider defaults
    for k, v := range p.options {
        if k != "system_prompt" {
            req[k] = v
        }
    }
    // merge per‐call overrides
    for k, v := range opts {
        if k != "system_prompt" {
            req[k] = v
        }
    }
    // apply any extra customizations
    for _, fn := range extras {
        fn(req)
    }
    return json.Marshal(req)
}

// Then simplify each Prepare… method. Example for PrepareRequest:
func (p *VLLMProvider) PrepareRequest(prompt string, options map[string]interface{}) ([]byte, error) {
    var msgs []map[string]interface{}
    if sp, ok := options["system_prompt"].(string); ok && sp != "" {
        msgs = append(msgs, map[string]interface{}{"role": "system", "content": sp})
    }
    msgs = append(msgs, map[string]interface{}{"role": "user", "content": prompt})
    return p.buildRequest(msgs, options)
}

// And for streaming:
func (p *VLLMProvider) PrepareStreamRequest(prompt string, options map[string]interface{}) ([]byte, error) {
    msgs := []map[string]interface{}{{"role": "user", "content": prompt}}
    return p.buildRequest(msgs, options, func(r map[string]interface{}) {
        r["stream"] = true
    })
}

// And for schema requests:
func (p *VLLMProvider) PrepareRequestWithSchema(
    prompt string, options map[string]interface{}, schemaObj interface{},
) ([]byte, error) {
    msgs := []map[string]interface{}{{"role":"user","content":prompt}}
    if sp, ok := options["system_prompt"].(string); ok && sp != "" {
        msgs = append([]map[string]interface{}{{"role":"system","content":sp}}, msgs...)
    }
    return p.buildRequest(msgs, options, func(r map[string]interface{}) {
        r["response_format"] = map[string]interface{}{"type": "json_object"}
        // if you need to inject the schema itself:
        // r["schema"] = schemaObj
    })
}
```
This removes the repeated message‐building and option‐merging logic while preserving every feature.
</issue_to_address>