new prompting mechanism and server services

data/bash.txt

@@ -0,0 +1,14 @@
+@role: user
+@content:
+<code>
+local utils = require("module.utils")
+local M = {}
+utils.pa<missing>
+return M
+</code>
+
+@role: assistant
+@content:
+<code>
+utils.parser
+</code>

lua/model_cmp/commands.lua

@@ -89,6 +89,7 @@ local function modelcmp_stop()
     vim.api.nvim_create_user_command('ModelCmpStop', function()
         common.stop()
         logger.debugging("Stopped api")
+        api.stop()
     end, {})
 end
 
@@ -103,6 +104,7 @@ local function modelcmp_logs()
         vim.api.nvim_buf_set_option(newbuf, 'swapfile', false)   -- No swap file
         vim.api.nvim_buf_set_lines(newbuf, 0, -1, false, logger.Logs)
         vim.api.nvim_buf_set_option(newbuf, 'modifiable', false) -- Make it read-only
+        -- Todo add logger
     end, {})
 end
 

lua/model_cmp/modelapi/common.lua

@@ -1,7 +1,7 @@
 local apiconfig = require("model_cmp.modelapi.apiconfig")
 local context = require("model_cmp.context")
 local logger = require("model_cmp.logger")
-local preprompt = require("model_cmp.modelapi.prompt")
+local prompter = require("model_cmp.modelapi.prompt")
 local req = require("model_cmp.modelapi.request")
 local utils = require("model_cmp.utils")
 local virtualtext = require("model_cmp.virtualtext")
@@ -11,7 +11,7 @@ local llama = require("model_cmp.modelapi.llama")
 
 local M = {}
 
-vim.g.server = "local_llama"
+vim.g.model_cmp_connection_server = nil
 
 local available_keys = {
     GEMINI_API_KEY = 0
@@ -35,25 +35,25 @@ local function check_available()
 end
 
 function M.send_request()
-    local bufnr = context.ContextEngine.bufnr
+    local currlang = context.ContextEngine.currlang
     local ctx = context.generate_context_text()
-    local few_shots = preprompt.complete_few_shots
+
+    local prompt = prompter.generate_prompt("text", ctx)
     local request
 
-    local server = vim.g.server
-    if vim.g.server == "" or vim.g.server == nil then
-        logger.error("NO server setup")
+    local server = vim.g.model_cmp_connection_server
+    if server == nil then
+        logger.trace("NO server setup")
         return
     end
-    if server == "default" or server == "local_llama" then
-        request = llama.generate_request(few_shots, ctx)
+    if server == "local_llama" then
+        request = llama.generate_request(prompt)
     elseif server == "gemini" then
         if available_keys.GEMINI_API_KEY ~= 1 then
             logger.error("GEMINI_API_KEY is not set")
             return
         end
-        local systemprompt = preprompt.default.content
-        request = gemini.generate_request(few_shots, systemprompt, ctx)
+        request = gemini.generate_request(prompt)
     end
 
     if request == nil then
@@ -63,8 +63,8 @@ function M.send_request()
     req.send(request,
         function(response)
             vim.schedule(function()
-                local text = ""
-                if vim.g.server == "gemini" then
+                local text = nil
+                if vim.g.model_cmp_connection_server == "gemini" then
                     text = gemini.decode_response(response)
                 else
                     text = utils.decode_response(response)
@@ -79,7 +79,7 @@ function M.send_request()
 end
 
 function M.stop()
-    vim.g.server = ""
+    vim.g.model_cmp_connection_server = nil
 end
 
 function M.setup(config)

lua/model_cmp/modelapi/gemini.lua

@@ -22,14 +22,15 @@ local function generate_url(model_name)
 end
 
 function M.start(model_name)
-    vim.g.server = "gemini"
+    vim.g.model_cmp_connection_server = "gemini"
 end
 
-local function transform_ctx_messages(ctx_messages)
-    -- Transforming few shot messages
+---@param prompt Prompt
+local function transform_fewshots(prompt)
     local new_chat = {}
-    for _, msg in ipairs(ctx_messages) do
+    for _, msg in ipairs(prompt.fewshots) do
         local gemini_message = {}
+
         if msg.role == 'user' then
             gemini_message = {
                 role = 'user',
@@ -45,17 +46,19 @@ local function transform_ctx_messages(ctx_messages)
                 },
             }
         end
+
         table.insert(new_chat, gemini_message)
     end
     return new_chat
 end
 
-function M.generate_request(ctx_messages, content, mainctx)
-    local messages = transform_ctx_messages(ctx_messages)
+---@param prompt Prompt
+function M.generate_request(prompt)
+    local messages = transform_fewshots(prompt)
     local mainmsg = {
         role = 'user',
         parts = {
-            { text = mainctx },
+            { text = prompt.context.content },
         },
     }
 
@@ -71,7 +74,7 @@ function M.generate_request(ctx_messages, content, mainctx)
         vim.fn.json_encode({
             system_instruction = {
                 parts = {
-                    text = content
+                    text = prompt.systemrole.content
                 }
             },
             contents = messages,

lua/model_cmp/modelapi/llama.lua

@@ -8,11 +8,23 @@ local generate_url = function(custom_url)
 end
 
 function M.start(model_name)
-    vim.g.server = ""
+    vim.g.model_cmp_connection_server = "local_llama"
 end
 
-function M.generate_request(ctx_messages, ctx)
+---@param prompt Prompt
+function M.generate_request(prompt)
     local custom = apiconfig.default()
+
+    local messages = { prompt.systemrole }
+    for _, k in ipairs(prompt.fewshots) do
+        table.insert(messages, k)
+    end
+    local context = {
+        role = "user",
+        content = prompt.context.content,
+    }
+    table.insert(messages, context)
+
     local request = {
         "-s",
         "-X", "POST",
@@ -21,7 +33,7 @@ function M.generate_request(ctx_messages, ctx)
         "-d",
         vim.fn.json_encode({
             model = "llama",
-            messages = ctx_messages,
+            messages = messages,
             n_predict = 128,
             temperature = 0.1,
             stop = { "</s>" },

lua/model_cmp/modelapi/prompt.lua

@@ -1,99 +1,77 @@
-local M = {}
-
-M.complete_few_shots = {
-    {
-        role = "user",
-        content = [[# language: lua
-<code>
-local ghosttext = require("model_cmp.ghosttext")
 local utils = require("model_cmp.utils")
 
 local M = {}
 
-vim.b.requ<missing>
+M.default_systemrole = {
+    role = "system",
+    content = [[Act as GitHub Copilot. Complete the code where the <missing> token is.
+Follow the instructions:
+- Output only the current line after replacing the <missing> tag.
+- No explanations, no comments, no full files generations allowed.
+- Max code generation is 5 lines.
+- Match the language and indentation.
+]]
+}
 
-local get_server_url = function() -- get server url from the user or from the config
-    return "http://127.0.0.1:8080/v1/chat/completions"
-end
-</code>]]
-    },
-    {
-        role = "assistant",
-        content = [[vim.b.request_sent = false]]
-    },
-    {
-        role = "user",
-        content = [[# language: lua
-<code>
-function M.get_trigger_characters()
-    return { '@', '.', '(', '[', ':', ' ' }
+---@param language string
+local function fewshot_lang_parser(language)
+    -- need to add path for the language context file
+    if language == "text" then
+        return {
+            {
+                role = "user",
+                content = "nothing"
+            },
+            {
+                role = "assistant",
+                content = "ok nothing"
+            }
+        }
+    end
 end
 
-function M.get_capabilities()
+---@param language string
+---@param ctx table<string>
+---@return Singlefewshot
+local function generate_context_shot(language, ctx)
+    local langprompt = "#language: " .. language
     return {
-        completion<missing>
-            triggerCharacters = M.get_trigger_characters(),
-        },
+        role = "user",
+        content = langprompt .. "\n" .. ctx
     }
 end
-</code>
-]]
-    },
-    {
-        role = "assistant",
-        content = [[        completionProvider = {]]
-    },
-    {
-        role = "user",
-        content = [[# language: Python
-<code>
-from transformers.agents import (
-    ReactCodeAgent,
-    ReactJsonAgent,
-    HfApiEngine,
-    ManagedAgent,
-)
-from transformers.agents.search import DuckDuckGoSearchTool
-
-llm_engine = <missing>
-
-web_agent = ReactJsonAgent(
-    tools=[DuckDuckGoSearchTool(), visit_webpage],
-    llm_engine=llm_engine,
-    max_iterations=10,
-)
-</code>
-]]
-    },
-    {
-        role = "assistant",
-        content = [[llm_engine = HfApiEngine(model)]]
-    },
-}
 
-M.default = {
-    role = "user",
-    content = [[Act as GitHub Copilot. Complete the code where the <missing> token is.
+---@class Singlefewshot
+---@field role string<"user" | "assistant" | "model" | "system">
+---@field content string
 
-Instructions:
-- Output only the code that replaces <missing>.
-- No explanations, no comments, no full file.
-- Limit to ≤ 2 lines of code.
-- Match language and indentation.
-]]
-}
+---@class Prompt
+---@field systemrole Singlefewshot
+---@field fewshots table<Singlefewshot>
+---@field language string
+---@field context Singlefewshot
 
--- After this we are going to collect and send new data
-M.closecall_suggestions = {
-    role = "user",
-    content = [[You almost got the right answer, try again with a different but similar result,
-]]
-}
+---@param ctx any
+---@return Prompt
+function M.default_prompt(ctx)
+    return {
+        systemrole = M.default_systemrole,
+        fewshots = fewshot_lang_parser("text"),
+        language = "text",
+        context = ctx
+    }
+end
 
-M.wrong_suggestion = {
-    role = "user",
-    content = [[You predicted wrong, try again with a completly new suggestion but with this code
-]]
-}
+function M.generate_prompt(language, ctx)
+    local prompt = M.default_prompt(ctx)
+    prompt.context = generate_context_shot(language, ctx)
+    if language == "text" or language == "" then
+        return prompt
+    end
+    local fewshots = fewshot_lang_parser(language)
+    prompt.fewshots = fewshots
+    prompt.language = language
+    return prompt
+end
 
 return M