Add R tool to build Gmail newsletter digests with local Ollama and README

NEWSLETTER_DIGEST_README.md

@@ -0,0 +1,48 @@
+# Gmail Newsletter Digest with Ollama (R version)
+
+This script pulls newsletter-like emails from Gmail and asks your local Ollama model to build a digest.
+
+## 1) Setup
+
+Install R packages:
+
+```r
+install.packages(c("argparse", "gmailr", "httr2", "jsonlite", "stringr"))
+```
+
+Enable Gmail API in Google Cloud and download OAuth Desktop credentials as `credentials.json` in this folder.
+
+## 2) Optional preference profile
+
+Create `preferences.txt` to guide ranking, for example:
+
+```text
+Prioritize AI engineering, data, Python, product strategy, and practical tutorials.
+Down-rank generic e-commerce promotions and celebrity/entertainment topics.
+```
+
+## 3) Run
+
+```bash
+Rscript gmail_newsletter_digest.R \
+  --query 'category:promotions newer_than:7d' \
+  --max-messages 40 \
+  --ollama-model llama3.1:8b \
+  --preferences-file preferences.txt \
+  --output newsletter_digest.md
+```
+
+Useful Gmail queries:
+- `label:newsletters newer_than:14d`
+- `category:promotions -from:(amazon.com) newer_than:7d`
+- `from:(substack.com OR beehiiv.com) newer_than:10d`
+
+## 4) Matching to your preferences (recommended approach)
+
+Use a lightweight feedback loop:
+1. Keep a `ratings.csv` with `message_id,rating` where rating is 1-5.
+2. Add those examples to the prompt each run so Ollama learns your taste.
+3. Promote senders/topics with avg rating >=4 and suppress <=2.
+4. Once enough data exists, train a tiny local classifier (logistic regression on subject + snippet embeddings) and combine with LLM score.
+
+This gives progressively better filtering while staying local/private.

gmail_newsletter_digest.R

@@ -0,0 +1,148 @@
+#!/usr/bin/env Rscript
+
+suppressPackageStartupMessages({
+  library(argparse)
+  library(gmailr)
+  library(httr2)
+  library(jsonlite)
+  library(stringr)
+})
+
+read_preferences <- function(path = NULL) {
+  if (is.null(path)) {
+    return(paste(
+      "Prefer newsletters about AI, software engineering, productivity, and practical tutorials.",
+      "Avoid celebrity news and generic promotions unless highly relevant."
+    ))
+  }
+  paste(readLines(path, warn = FALSE, encoding = "UTF-8"), collapse = "\n")
+}
+
+extract_header <- function(headers, header_name) {
+  matches <- Filter(function(h) identical(tolower(h$name), tolower(header_name)), headers)
+  if (length(matches) == 0) return("")
+  matches[[1]]$value %||% ""
+}
+
+`%||%` <- function(a, b) if (is.null(a)) b else a
+
+extract_text <- function(payload) {
+  if (!is.null(payload$parts)) {
+    plain_part <- NULL
+    html_part <- NULL
+
+    for (part in payload$parts) {
+      mime <- part$mimeType %||% ""
+      if (identical(mime, "text/plain") && !is.null(part$body$data)) {
+        plain_part <- part
+        break
+      }
+      if (identical(mime, "text/html") && !is.null(part$body$data)) {
+        html_part <- part
+      }
+    }
+
+    target <- plain_part %||% html_part
+    if (is.null(target)) return("")
+
+    raw <- base64urldecode(target$body$data)
+    txt <- rawToChar(raw)
+    if (identical(target$mimeType, "text/html")) {
+      txt <- str_replace_all(txt, "<[^>]+>", " ")
+      txt <- str_squish(txt)
+    }
+    return(txt)
+  }
+
+  if (!is.null(payload$body$data)) {
+    return(rawToChar(base64urldecode(payload$body$data)))
+  }
+
+  ""
+}
+
+fetch_newsletters <- function(query, max_messages) {
+  ids <- gm_messages(search = query, num_results = max_messages)
+  if (length(ids) == 0) return(list())
+
+  out <- list()
+  for (id in ids$id) {
+    msg <- gm_message(id)
+    payload <- msg$payload %||% list()
+    headers <- payload$headers %||% list()
+
+    item <- list(
+      message_id = id,
+      from = extract_header(headers, "From"),
+      subject = extract_header(headers, "Subject"),
+      date = extract_header(headers, "Date"),
+      snippet = msg$snippet %||% "",
+      content = str_sub(extract_text(payload), 1, 8000)
+    )
+    out[[length(out) + 1]] <- item
+  }
+  out
+}
+
+build_prompt <- function(messages, preferences) {
+  serialized <- toJSON(messages, auto_unbox = TRUE, pretty = FALSE, null = "null")
+  paste0(
+    "You are a newsletter assistant. Given emails and user preferences, produce:\n",
+    "1) A concise digest grouped by theme.\n",
+    "2) Top 10 most relevant newsletters with one-line rationale.\n",
+    "3) A scoring rubric (0-100) based on user preferences.\n",
+    "4) A plan to improve matching over time.\n\n",
+    "User preferences:\n", preferences, "\n\n",
+    "Emails JSON:\n", str_sub(serialized, 1, 120000)
+  )
+}
+
+call_ollama <- function(model, host, prompt) {
+  req <- request(host) |>
+    req_url_path_append("api", "generate") |>
+    req_body_json(list(model = model, prompt = prompt, stream = FALSE)) |>
+    req_timeout(120)
+
+  resp <- req_perform(req)
+  body <- resp_body_json(resp, simplifyVector = TRUE)
+  body$response %||% ""
+}
+
+write_digest <- function(path, content) {
+  writeLines(enc2utf8(content), con = path, useBytes = TRUE)
+}
+
+parse_args <- function() {
+  parser <- ArgumentParser(description = "Create a Gmail newsletter digest with local Ollama")
+  parser$add_argument("--query", default = "category:promotions newer_than:7d")
+  parser$add_argument("--max-messages", type = "integer", default = 30)
+  parser$add_argument("--ollama-model", default = "llama3.1:8b")
+  parser$add_argument("--ollama-host", default = "http://127.0.0.1:11434")
+  parser$add_argument("--preferences-file", default = NULL)
+  parser$add_argument("--output", default = "newsletter_digest.md")
+  parser$parse_args()
+}
+
+main <- function() {
+  args <- parse_args()
+
+  gm_auth_configure(path = "credentials.json")
+  gm_auth(email = TRUE)
+
+  preferences <- read_preferences(args$`preferences-file`)
+  messages <- fetch_newsletters(args$query, args$`max-messages`)
+
+  if (length(messages) == 0) {
+    write_digest(args$output, "No messages found for this query.")
+    cat("Digest created:", args$output, "\n")
+    return(invisible(NULL))
+  }
+
+  prompt <- build_prompt(messages, preferences)
+  digest <- call_ollama(args$`ollama-model`, args$`ollama-host`, prompt)
+  write_digest(args$output, digest)
+
+  cat("Digest created:", args$output, "\n")
+}
+
+main()