Adds a minimum word size to send context summary

DESCRIPTION

@@ -1,6 +1,6 @@
 Package: lang
 Title: Translates R Help Documentation using Large Language Models
-Version: 0.1.1
+Version: 0.1.1.9000
 Authors@R: c(
     person("Edgar", "Ruiz", , "edgar@posit.co", role = c("aut", "cre")),
     person("Posit Software, PBC", role = c("cph", "fnd"),

NEWS.md

@@ -1,3 +1,9 @@
+# lang 0.1.1.9000
+
+## Bug fixes
+
+* When the input is 10 words or fewer, the context summary is now omitted from the translation prompt. Local LLMs can get confused by a context summary that is much longer than the field being translated, causing them to paraphrase the context instead of translating the input.
+
 # lang 0.1.1
 
 ## New features

R/lang-help.R

@@ -8,6 +8,14 @@
 #' will be processed, so the help returned will be the original package's
 #' documentation.
 #'
+#' To improve translation consistency, `lang` generates a short summary of the
+#' full help page and passes it as context with each individual translation
+#' request. This helps the LLM maintain consistent terminology across sections.
+#' You can control the length of this summary with the `context_size` argument,
+#' or set it to `0` to disable context-aware translation entirely. To avoid the
+#' LLM getting confused when the context is much longer than the field being
+#' translated, context is automatically omitted for fields of 10 words or fewer.
+#'
 #' @param topic A character string specifying the help topic to translate.
 #' @param package The R package to look for the topic, if not provided the
 #' function will attempt to find the topic based on the loaded packages.

R/rd-translate.R

@@ -173,7 +173,9 @@ rd_translate <- function(rd_content, lang, context_size) {
 }
 
 rd_field_translate <- function(x, lang, rs, context_summary = NULL) {
-  context_block <- if (!is.null(context_summary)) {
+  word_count <- length(unlist(strsplit(paste(x, collapse = " "), " ")))
+  use_context <- !is.null(context_summary) && word_count > 10L
+  context_block <- if (use_context) {
     paste0(
       "For context, here is a short summary of the full help page in the target language:\n\n",
       context_summary,
@@ -288,6 +290,12 @@ lang_rs_hash <- function() {
 }
 
 lang_rs_get <- function() {
+  if (is.null(.lang_env$session[["backend"]])) {
+    cli_abort(
+      "No LLM backend configured. Call {.fn lang_use} first.",
+      call = NULL
+    )
+  }
   rs <- .lang_env$rs
   if (!is.null(rs) && rs$is_alive()) {
     if (rs$get_state() == "starting") {

README.Rmd

@@ -76,16 +76,24 @@ Note that R enforces the printed names of each section, so titles such as
 "Description", "Usage", and "Arguments" will always remain untranslated.
 
 During translation, `lang` will display its progress by showing which section
-of the documentation is currently translating. Because each section of a help
-page is translated independently, the LLM can lose track of the broader topic
-and produce inconsistent or out-of-context translations. To address this,
-`lang` first summarizes the full help page in English, translates that summary
-into the target language, and then uses it as context when translating each
-individual section. You can control the length of this summary with the
-`context_size` argument in `lang_use()` or `lang_help()` — set it to `0` to
-disable it, or increase it to give the LLM more context. During the R session,
-if you request the same R function's help more than one time then `lang` will
-use its cached results, which will run immediately.
+of the documentation is currently translating. During the R session, if you
+request the same R function's help more than one time then `lang` will use its
+cached results, which will run immediately.
+
+### Context summary
+
+Because each section of a help page is translated independently, the LLM can
+lose track of the broader topic and produce inconsistent or out-of-context
+translations. To address this, `lang` first summarizes the full help page in
+English, translates that summary into the target language, and then uses it as
+context when translating each individual section. You can control the length of
+this summary with the `context_size` argument in `lang_use()` or `lang_help()`
+— set it to `0` to disable it, or increase it to give the LLM more context.
+
+To avoid the LLM getting confused by a context summary that is longer than the
+content being translated, context is automatically omitted for fields of 10
+words or fewer.
+
 
 
 ### LLM connections
@@ -228,6 +236,14 @@ use the full language name, such as 'spanish', or 'french', etc.  You can use
 `Sys.setenv(LANGUAGE = "[my language]")`, or, for a more permanent solution, 
 add the entry to your .Renviron file (`usethis::edit_r_environ()`).
 
+### Translation errors
+
+If you experience unexpected translation errors and you are using a local LLM
+without a `seed` set, try restarting your R session and running the translation
+again. Non-deterministic LLM output can occasionally produce output that causes
+errors. If the problem persists, please open an issue at
+<https://github.com/mlverse/lang/issues>.
+
 ### Interaction with `mall`
 
 `lang` uses the `mall` package to produce the translations. To avoid conflicts