fix formatting output synopses_generator

closes #4

ReadMe.md

@@ -19,19 +19,22 @@ Opens `similarities.json`, reverses the sort order, and saves it as
 
 ### `refresh_sources.lua`
 Creates/Maintains a store of chunked data with embeddings based on configurable
-sources. Example:
+`sources.json`. Example:
 
 ```json
 {
   "source name":{
+    "embedding_model":"qwen3-embedding:0.6b",
     "filters":{
       "blacklist":[".git"],
       "extension_whitelist":["md"]
     },
     "initialize_command":"git clone REMOTE .",
+    "max_chunk_size":32768,
     "path":"will be created before initialize_command is run",
     "strip_frontmatter":true,
-    "refresh_command":"git reset --hard origin/main"
+    "target_chunk_size":3072,
+    "refresh_command":"git fetch origin && git reset --hard origin/main"
   }
 }
 ```
@@ -41,6 +44,9 @@ The filters are based on `utility.tree`'s filter options (optional).
 while `refresh_command` is run each time (optional).
 Commands will be run in the specified `path`.
 `strip_frontmatter` will remove YAML frontmatter (common in Markdown files).
+`embedding_model` and `max_chunk_size` are self-explanatory (optional).
+`target_chunk_size` (optional) allows for finer-grained chunks to be generated
+alongside the maximum overviews for more precise retrieval.
 
 The embeddings are stored like so:
 
@@ -55,7 +61,11 @@ The embeddings are stored like so:
 }
 ```
 
-TODO
+SHA2 512-bit sums are used to link a text chunk with its embedding. All file
+names reference a list of text chunks so they can handle being too large. Every
+file that is too large for a single chunk has a whole-file embedding calculated
+first (it is the first element of the array), so that if a whole file becomes
+relevant, it can still show up instead of only chunks.
 
 ### `synopsis_generator.lua`
 Chooses a random file within `notebook`, and generates a novel synopsis from it.
@@ -64,6 +74,28 @@ Arguments:
 - `refresh_file_list`: Refreshes the cached file list to choose from.
 - `export_ordered_list_of_prompts`: Makes an epub to review generated synopses.
 
+## JSON config
+This repo uses my utility library's config system, using a `config.json` file in
+the repo root that is excluded from commits.
+
+```json
+{
+  "models":{
+    "embedding":{
+      "model":"qwen3-embedding:0.6b",
+      "max_chunk_size":32768
+    },
+    "initialized_sources":{}
+  }
+}
+```
+
+`refresh_sources.lua` uses `models` to store default embedding model information
+and `initialized_sources` to store which sources have been initialized.
+
 ## Tasks
 - [ ] The whitelisting/blacklisting of tree should be in list too.
 - [ ] synopsis_generator should be able to blacklist files it already tried?
+- [ ] `refresh_sources.lua` doesn't check for defined files that don't exist
+  anymore, does it?
+  - [ ] I think it checks for every other possibility, but this needs checking.

lib/log.lua

@@ -1,11 +1,8 @@
-local stored_messages = {}
-local display_levels = {}
-local enable_logging = true
-local message_limit = math.huge
-
 -- USAGE:
 --   -- set arbitrary log levels to be printed immediately
 --   log{ info = true, warning = true, ducksauce = true, }
+--   -- disable storage of messages depending on log level
+--   log{ error = false, verbose = false, }
 --   -- send anything (except nil) to a log level
 --   log("bacom", true, "text", 5, function() end, {})
 --   -- print all messages saved at a particular log level
@@ -14,62 +11,72 @@ local message_limit = math.huge
 --   log(false) -- this deletes all stored messages
 --   -- set a maximum number of stored messages per log level
 --   log(50) -- warning: reaching large limits makes logging expensive
+--  -- check if a log level is being displayed
+--  if log.arbitrary_exit then log("arbitrary_exit") os.exit(1) end
 --   -- do anything to the stored messages (this example deletes them all)
 --  log(function(messages) return {} end)
 
-return function(options, ...)
-  local options_type = type(options)
+local microlog = {} -- display_levels are stored directly here
+local stored_messages = {}
+local enable_logging = true
+local message_limit = math.huge
 
-  if (options_type == "string") and enable_logging then
-    if not stored_messages[options] then stored_messages[options] = {} end
-    local message_table = stored_messages[options]
+return setmetatable(microlog, {
+  __call = function(self, options, ...)
+    local options_type = type(options)
 
-    -- turn log message into text
-    local tab = {...}
-    for i = 1, #tab do
-      tab[i] = tostring(tab[i])
-    end
-    local current_message = table.concat(tab, "\t")
+    if (options_type == "string") and enable_logging then
+      if microlog[options] == false then return end
+      if not stored_messages[options] then stored_messages[options] = {} end
+      local message_table = stored_messages[options]
 
-    if #current_message == 0 then
-      -- print all stored_messages of specified level
-      for i = 1, #message_table do
-        print(message_table[i])
+      -- turn log message into text
+      local tab = {...}
+      for i = 1, #tab do
+        tab[i] = tostring(tab[i])
+      end
+      local current_message = table.concat(tab, "\t")
+
+      if #current_message == 0 then
+        -- print all stored_messages of specified level
+        for i = 1, #message_table do
+          print(message_table[i])
+        end
+
+      else
+        -- store message (and print if in display_levels)
+        message_table[#message_table + 1] = current_message
+        if #message_table > message_limit then table.remove(message_table, 1) end
+        if microlog[options] then print(current_message) end
       end
 
-    else
-      -- store message (and print if in display_levels)
-      message_table[#message_table + 1] = current_message
-      if #message_table > message_limit then table.remove(message_table, 1) end
-      if display_levels[options] then print(current_message) end
-    end
-
-  -- set display_levels
-  elseif options_type == "table" then
-    for k,v in pairs(options) do
-      display_levels[k] = v
-    end
-
-  elseif options_type == "number" then
-    message_limit = options
-    for k,v in pairs(stored_messages) do
-      while #v > message_limit do
-        table.remove(v, 1)
+    -- set display_levels
+    elseif options_type == "table" then
+      for k,v in pairs(options) do
+        microlog[k] = v
       end
-    end
 
-  -- turn all logging off and delete stored messages; or turn it on
-  elseif options_type == "boolean" then
-    if options then
-      enable_logging = true
-    else
-      enable_logging = false
-      stored_messages = {}
-    end
+    elseif options_type == "number" then
+      message_limit = options
+      for k,v in pairs(stored_messages) do
+        while #v > message_limit do
+          table.remove(v, 1)
+        end
+      end
 
-  -- arbitrary access :D
-  elseif options_type == "function" then
-    local result = options(stored_messages)
-    if type(result) == "table" then stored_messages = result end
+    -- turn all logging off and delete stored messages; or turn it on
+    elseif options_type == "boolean" then
+      if options then
+        enable_logging = true
+      else
+        enable_logging = false
+        stored_messages = {}
+      end
+
+    -- arbitrary access :D
+    elseif options_type == "function" then
+      local result = options(stored_messages)
+      if type(result) == "table" then stored_messages = result end
+    end
   end
-end
+})

lib/utility.lua

@@ -375,10 +375,10 @@ end
 
 
 
+local config_path = utility.path .. "config.json"
 local config, config_lock
 utility.get_config = function(skip_lock)
   if not config then
-    local config_path = utility.path .. "config.json"
     if utility.is_file(config_path) then
       if not skip_lock then
         config_lock = utility.get_lock(config_path)
@@ -396,7 +396,6 @@ end
 
 utility.save_config = function()
   if config then
-    local config_path = utility.path .. "config.json"
     if not config_lock then
       print("Warning: A config lock file was not established.")
     end
@@ -412,6 +411,31 @@ utility.save_config = function()
   end
 end
 
+utility.get_config_with_defaults = function(defaults)
+  local config = utility.get_config()
+
+  local loop, changes_made
+  loop = function(config_level, default_level)
+    for k,v in pairs(default_level) do
+      if not config_level[k] then
+        config_level[k] = v
+        changes_made = true
+      elseif type(v) == "table" then
+        loop(config_level[k], v)
+      end
+    end
+  end
+  loop(config, defaults)
+
+  if changes_made then
+    utility.save_config()
+  else
+    utility.release_lock(config_path)
+  end
+
+  return config
+end
+
 
 
 local data_file_locations = {}

list_file_extensions.lua

@@ -0,0 +1,19 @@
+#!/usr/bin/env luajit
+
+package.path = (arg[0]:match("@?(.*/)") or arg[0]:match("@?(.*\\)")) .. "lib" .. package.config:sub(1, 1) .. "?.lua;" .. package.path
+local utility = require "utility"
+local json = utility.require("dkjson")
+
+local search_path = arg[1]
+
+local file_extensions = {}
+utility.tree(search_path, function(file_path)
+  local _, _, extension = utility.split_path_components(file_path)
+  if extension then
+    file_extensions[extension] = true
+  end
+end)
+
+for extension in pairs(file_extensions) do
+  print(extension)
+end

refresh_sources.lua

@@ -8,24 +8,21 @@ local log = utility.require("log")
 local text_processing = utility.require("text_processing")
 local timing = utility.require("timing")
 
--- TODO make utility have a function for getting/setting defaults where locking is only used to set defaults if they aren't present
 -- TODO there should be a check function for a lock so a warning/error can be dumped?
-local config = utility.get_config("no-lock")
-if not config.models then
-  config.models = {
+local config = utility.get_config_with_defaults{
+  models = {
     embedding = {
       model = "qwen3-embedding:0.6b",
       max_chunk_size = 32768,
     },
     initialized_sources = {},
-  }
-  utility.save_config()
-end
+  },
+}
 
 log{
   info = true,
   warning = true,
-  -- debug = true,
+  debug = false,
   -- files = true, -- debugging why the wrong files are selected
   -- sha = true, -- what the fuck is going on with sha sums?
 }
@@ -34,7 +31,6 @@ local memory_path = "PRIVATE_DATA" .. utility.path_separator .. "memory"
 local embeddings_file_path = memory_path .. utility.path_separator .. "+embeddings.json"
 local tmp_file_path = "PRIVATE_DATA" .. utility.path_separator .. ".tmp.2b65c19b-0883-49ca-8247-b1fe7760f922"
 
-local embeddings
 if not utility.path_exists(embeddings_file_path) then
   os.execute("mkdir -p " .. memory_path:enquote())
   utility.save_data({
@@ -42,8 +38,8 @@ if not utility.path_exists(embeddings_file_path) then
     vectors = {},
   }, embeddings_file_path)
 end
-embeddings = utility.load_data(embeddings_file_path)
-local function embeddings_debug()
+local embeddings = utility.load_data(embeddings_file_path)
+if log.debug then
   log("debug", "Embeddings loaded.", embeddings, embeddings.files, embeddings.vectors)
   local file_count = 0
   for k,v in pairs(embeddings.files) do
@@ -57,12 +53,11 @@ local function embeddings_debug()
   log("debug", vector_count .. " vectors.")
   -- os.exit(1)
 end
-embeddings_debug()
 
 
 
 local refresh_file_list = function(source_name, data_source)
-  timing.mark("Assembling file list for " .. source_name .. ".")
+  timing.mark("Assembling file list for \"" .. source_name .. "\"")
 
   local full_path = "PRIVATE_DATA" .. utility.path_separator .. data_source.path
   if data_source.initialize_command and (not config.models.initialized_sources[data_source.path]) then
@@ -88,34 +83,26 @@ local refresh_file_list = function(source_name, data_source)
     file_list[#file_list + 1] = file_name
   end)
 
-  timing.mark("Finished assembling file list for " .. source_name .. ".")
+  timing.mark("Finished assembling file list for \"" .. source_name .. "\"")
   return file_list
 end
 
 -- returns nothing when too much text is sent
-local generate_embeddings = function(text)
-  if #text > config.models.embedding.max_chunk_size then
+local generate_embeddings = function(data_source, text)
+  local max_chunk_size = data_source.max_chunk_size or config.models.embedding.max_chunk_size
+  local model = data_source.embedding_model or config.models.embedding.model
+
+  if #text > max_chunk_size then
     return nil, "generate_embeddings() must only be passed appropriately-sized chunks!"
   end
 
-  local result = utility.llm_prompt(text, config.models.embedding.model)
+  local result = utility.llm_prompt(text, model)
   return json.decode(result)
 end
 
--- returns nothing for empty files and errors
-local process_file = function(data_source, file_name)
-  local text = utility.read_file(file_name)
+local make_chunks = function(text, chunk_size)
+  assert(chunk_size, "make_chunks() requires chunk_size")
 
-  if data_source.strip_frontmatter then
-    text = text_processing.strip_frontmatter(text)
-  end
-
-  if #text == 0 then
-    log("empty", file_name .. "\n is empty and being skipped.")
-    return
-  end
-
-  local chunk_size = config.models.embedding.max_chunk_size
   local half_chunk_size = math.floor(chunk_size / 2)
   local chunks = { text }
 
@@ -133,16 +120,44 @@ local process_file = function(data_source, file_name)
     end
   end
 
+  return chunks
+end
+
+-- returns nothing for empty files and errors
+local process_file = function(data_source, file_name)
+  local text = utility.read_file(file_name)
+
+  if data_source.strip_frontmatter then
+    text = text_processing.strip_frontmatter(text)
+  end
+
+  if #text == 0 then
+    log("empty", file_name .. "\n is empty and being skipped.")
+    return
+  end
+
+  local chunk_size = data_source.max_chunk_size or config.models.embedding.max_chunk_size
+  local chunks = make_chunks(text, chunk_size)
+
+  if data_source.target_chunk_size then
+    local extra_chunks = make_chunks(text, data_source.target_chunk_size)
+    for i = 2, #extra_chunks do
+      chunks[#chunks + 1] = extra_chunks[i]
+    end
+  end
+
   local new_embeddings = {}
   for i = 1, #chunks do
-    log("debug", "Embedding length:", #chunks[i])
-    new_embeddings[i] = generate_embeddings(chunks[i]) or {}
+    log("debug", "Embedding source length:", #chunks[i])
+    new_embeddings[i] = generate_embeddings(data_source, chunks[i]) or {}
   end
 
   if #new_embeddings[1] == 0 then
-    log("debug", "Vector lengths:")
-    for e = 1, #new_embeddings do
-      log("debug", "", e, #new_embeddings[e])
+    if log.debug then
+      log("debug", "Vector lengths:")
+      for e = 1, #new_embeddings do
+        log("debug", "", e, #new_embeddings[e])
+      end
     end
     if #new_embeddings == 1 then
       -- Ollama very rarely errors with:
@@ -198,6 +213,15 @@ local memorize_file = function(data_source, file_name)
   return file_sums
 end
 
+local delete_memories = function(sha512sum_list)
+  -- if true then return end -- TEMP disabling removal to compare and verify function
+  for i = 1, #sha512sum_list do
+    local sha512sum = sha512sum_list[i]
+    embeddings.vectors[sha512sum] = nil
+    os.execute("rm " .. (memory_path .. utility.path_separator .. sha512sum):enquote())
+  end
+end
+
 local refresh_sources = function()
   os.execute("mkdir -p " .. memory_path:enquote())
 
@@ -226,16 +250,23 @@ local refresh_sources = function()
       local function loop()
         local sha512sum = utility.sha512sum(file_name)
         log("sha", file_name, sha512sum, "\n Sum present? " .. tostring(embeddings.vectors[sha512sum]))
+
         if embeddings.vectors[sha512sum] then
           log("debug", file_name .. "\n has already been embedded, skipping.")
-          -- add file reference if it was missing
+          -- add file reference if it was missing (also handles recognition of duplicate files)
           if not embeddings.files[file_name] then
             embeddings.files[file_name] = { sha512sum }
           end
           return
         end
 
-        log("debug", "Generating embeddings for " .. file_name .. ".")
+        if embeddings.files[file_name] then
+          log("debug", file_name .. "\n Removing old embeddings for modified file.")
+          delete_memories(embeddings.files[file_name])
+          embeddings.files[file_name] = nil
+        end
+
+        log("debug", file_name .. "\n Generating embeddings..")
         local file_sums = memorize_file(data_source, file_name)
         embeddings.files[file_name] = file_sums
       end

synopsis_generator.lua

@@ -7,49 +7,49 @@ local json = utility.require("dkjson")
 local prompts = utility.require("prompts")
 local text_processing = utility.require("text_processing")
 
-local model = "gemma4:12b-mlx"
-local minimum_bytes = 1000
-local maximum_bytes = 40000
+local config = utility.get_config_with_defaults{
+  models = {
+    synopsis_generator = {
+      model = "gemma4:12b-mlx",
+      minimum_bytes = 1000,
+      maximum_bytes = 40000,
+    },
+  },
+}
 
-local NOTEBOOK_PATH = "PRIVATE_DATA/notebook"
-local blacklist = utility.enumerate{ ".git", ".gitattributes", ".gitignore", ".gitkeep", ".DS_Store", }
-local extension_blacklist = utility.enumerate{ "gif", "jpg", "jpeg", "mp4", "pdf", "png", "webp", }
-
-local data_location = "PRIVATE_DATA/synopsis_generator_file_list.json"
-local file_list
-if utility.path_exists(data_location) then
-  file_list = utility.load_data(data_location)
-else
-  file_list = {}
-end
+local memory_path = "PRIVATE_DATA" .. utility.path_separator .. "memory"
+local embeddings_file_path = memory_path .. utility.path_separator .. "+embeddings.json"
+local data_path = "PRIVATE_DATA" .. utility.path_separator .. "synopses"
+local data_location = data_path .. utility.path_separator .. "+file_list.json"
 
 local refresh_file_list = function()
-  os.execute("cd " .. NOTEBOOK_PATH:enquot() .. " && git pull origin")
+  assert(utility.path_exists(embeddings_file_path), "Run \"./refresh_sources.lua\" to set up memory.")
+  local embeddings = utility.load_data(embeddings_file_path)
 
-  local new_files_list = {}
-  utility.tree(NOTEBOOK_PATH, {
-    blacklist = blacklist,
-    extension_blacklist = extension_blacklist,
-  }, function(file_name)
+  local minimum_bytes = config.models.synopsis_generator.minimum_bytes
+  local maximum_bytes = config.models.synopsis_generator.maximum_bytes
+
+  local file_list = {}
+  for file_name, sha512sum_list in pairs(embeddings.files) do
     local file_size = utility.file_size(file_name)
     if file_size >= minimum_bytes and file_size <= maximum_bytes then
       local text = text_processing.strip_frontmatter(utility.read_file(file_name))
       if #text >= minimum_bytes and #text <= maximum_bytes then
-        new_files_list[#new_files_list + 1] = file_name
+        file_list[#file_list + 1] = { file_name = file_name, }
       end
     end
-  end)
+  end
 
-  file_list = new_files_list
-  utility.save_data(new_files_list, data_location)
+  utility.save_data(file_list, data_location)
+  return file_list
 end
 
 local generate_and_score = function(file_name, text)
   print("Writing synopsis...")
-  local synopsis = utility.llm_prompt(prompts.synopsis_prompt .. text, model)
+  local synopsis = utility.llm_prompt(prompts.synopsis_prompt .. text, config.models.synopsis_generator.model)
   print(synopsis)
   print("Scoring synopsis...")
-  local scoring, thinking = utility.llm_prompt(prompts.scoring_prompt .. synopsis, model)
+  local scoring, thinking = utility.llm_prompt(prompts.scoring_prompt .. synopsis, config.models.synopsis_generator.model)
   print(scoring)
   local scoring_decoded = json.decode(scoring) -- likely will not work because it consistently returns Markdown instead of JSON
   if not scoring_decoded then
@@ -62,23 +62,22 @@ local generate_and_score = function(file_name, text)
     thinking = thinking,
   }
 
-  utility.save_data(object, "PRIVATE_DATA/synopses/" .. utility.uuid() .. ".json")
+  utility.save_data(object, data_path .. utility.path_separator .. utility.uuid() .. ".json")
 end
 
 local export_ordered_list_of_prompts = function()
-  local path = "PRIVATE_DATA/synopses"
   local items = {}
   local item_order = {}
 
-  utility.list(path, function(path_name)
-    local full_path = path .. utility.path_separator .. path_name
-    if path_name:find("%.json") then
-      local object = utility.load_data(full_path)
-      items[path_name] = object
-      if type(object.scoring) == "table" then
-        local mean_score, total_score = utility.mean(object.scoring)
-        item_order[#item_order + 1] = { path_name = path_name, total_score = total_score, mean_score = mean_score, }
-      end
+  utility.list(data_path, function(path_name)
+    local full_path = data_path .. utility.path_separator .. path_name
+    if full_path == data_location then return end
+
+    local object = utility.load_data(full_path)
+    items[path_name] = object
+    if type(object.scoring) == "table" then
+      local mean_score, total_score = utility.mean(object.scoring)
+      item_order[#item_order + 1] = { path_name = path_name, total_score = total_score, mean_score = mean_score, }
     end
   end)
 
@@ -87,7 +86,7 @@ local export_ordered_list_of_prompts = function()
   local output = {
     "---",
     "title: Ordered Synopses (" .. #item_order .. " items)",
-    "author: [\"" .. model .. "\", \"Tangent\", \"Ollama\"]",
+    "author: [\"" .. config.models.synopsis_generator.model .. "\", \"Tangent\", \"Ollama\"]",
     "publisher: \"synopsis_generator.lua\"",
     "---",
     "",
@@ -99,8 +98,11 @@ local export_ordered_list_of_prompts = function()
     local tab = text:split("\n")
 
     for index, line in ipairs(tab) do
-      if line:sub(1, 1) == "#" then
-        tab[index] = "#" .. tab[index]
+      -- ensure no output heading is an H1; make H6 bold text lines instead
+      if line:sub(1, 7) == "###### " then
+        tab[index] = "**" .. line:sub(8) .. "**"
+      elseif line:sub(1, 1) == "#" then
+        tab[index] = "#" .. line
       end
     end
 
@@ -114,25 +116,27 @@ end
 
 
 
-os.execute("mkdir -p PRIVATE_DATA/synopses")
+os.execute("mkdir -p " .. data_path:enquote())
+
+local file_list
+if utility.path_exists(data_location) then
+  file_list = utility.load_data(data_location)
+end
 
 if arg[1] == "refresh_file_list" then
   print("Refresing file list...")
-  refresh_file_list()
+  file_list = refresh_file_list()
   os.exit(0)
 elseif arg[1] == "export_ordered_list_of_prompts" then
   export_ordered_list_of_prompts()
   os.exit(0)
 end
 
-print(#file_list .. " files to select from.")
-if #file_list == 0 then
-  print("Run \"./synopsis_generator.lua refresh_file_list\" first.")
-  os.exit(1)
-end
+assert(file_list and (#file_list > 0), "Run \"./synopsis_generator.lua refresh_file_list\" first.")
 
+print(#file_list .. " files to select from.")
 while true do
-  local file_name = file_list[math.random(1, #file_list)]
+  local file_name = file_list[math.random(1, #file_list)].file_name
   local text = utility.read_file(file_name)
   print(file_name .. " chosen.")
   generate_and_score(file_name, text)