branch: externals/minuet
commit cf16b7b2859fabae8934f28d3cafb83c6f9f72a4
Author: Milan Glacier <d...@milanglacier.com>
Commit: Milan Glacier <d...@milanglacier.com>
feat!: Improve completion filtering with before/after context.
This commit refactors the completion filtering logic to based on
longest common match and add support for filtering based on prefix.
Key changes:
- Replaces the simple substring removal with a more precise
`minuet-find-longest-match` function, which finds the longest common
substring between a completion candidate and the surrounding context.
- Introduces `minuet-before-cursor-filter-length` to trim redundant
prefixes from completions based on the text before the cursor.
---
minuet.el | 129 ++++++++++++++++++++++++++++++++++++++++++++------------------
1 file changed, 91 insertions(+), 38 deletions(-)
diff --git a/minuet.el b/minuet.el
index b53fd07f91..4f10f98c72 100644
--- a/minuet.el
+++ b/minuet.el
@@ -152,15 +152,37 @@ complete error log."
(defcustom minuet-after-cursor-filter-length 15
"Length of context after cursor used to filter completion text.
-Defines the length of non-whitespace context after the cursor used to
-filter completion text. Set to 0 to disable filtering.
-
-Example: With after_cursor_filter_length = 3 and context: \"def
-fib(n):\\n|\\n\\nfib(5)\" (where | represents cursor position), if the
-completion text contains \"fib\", then \"fib\" and subsequent text
-will be removed. This setting filters repeated text generated by the
-LLM. A large value (e.g., 15) is recommended to avoid false
-positives."
+This setting helps prevent the language model from generating
+redundant text. When filtering completions, the system compares the
+suffix of a completion candidate with the text immediately following
+the cursor.
+
+If the length of the longest common substring between the end of the
+candidate and the beginning of the post-cursor context exceeds this
+value, that common portion is trimmed from the candidate.
+
+For example, if the value is 15, and a completion candidate ends with
+a 20-character string that exactly matches the 20 characters following
+the cursor, the candidate will be truncated by those 20 characters
+before being presented."
+ :type 'integer)
+
+(defcustom minuet-before-cursor-filter-length 3
+ "Length of context before cursor used to filter completion text.
+
+This setting helps prevent the language model from generating
+redundant text at the beginning of completion. When filtering
+completions, the system compares the prefix of a completion candidate
+with the text immediately before the cursor.
+
+If the length of the longest common substring between the beginning of
+the candidate and the end of the pre-cursor context exceeds this
+value, that common portion is trimmed from the candidate.
+
+For example, if the value is 3, and a completion candidate starts with
+a 10-character string where the last 3 characters exactly match the 3
+characters before the cursor, the candidate will be truncated by those
+3 characters before being presented."
:type 'integer)
(defcustom minuet-n-completions 3
@@ -688,40 +710,71 @@ conversation with alternating `user` and `assistant`
roles by
(push (apply #'concat (nreverse parts)) results))
(nreverse results)))
-(defun minuet--make-context-filter-sequence (context len)
- "Create a filtering string based on CONTEXT with maximum length LEN."
- (if-let* ((is-string (stringp context))
- (is-positive (> len 0))
- (context (replace-regexp-in-string "\\`[\s\t\n]+" "" context))
- (should-filter (>= (length context) len))
- (context (substring context 0 len))
- (context (replace-regexp-in-string "[\s\t\n]+\\'" "" context)))
- context
+(cl-defun minuet--filter-text (item context)
+ "Filter ITEM based on CONTEXT using minuet-find-longest-match.
+ITEM is a completion candidate string.
+CONTEXT is a plist with :before-cursor and :after-cursor fields.
+Returns the filtered item after trimming overlapping parts."
+ (when (null item)
+ (cl-return-from minuet--filter-text nil))
+
+ (when (null context)
+ (cl-return-from minuet--filter-text item))
+
+ (setq item (string-trim item))
+
+ (let* ((before-cursor (plist-get context :before-cursor))
+ (after-cursor (plist-get context :after-cursor))
+ (filtered-item item))
+
+ ;; Filter against before-cursor context (trim from prefix)
+ (when (and before-cursor
+ (> minuet-before-cursor-filter-length 0))
+ (setq before-cursor (string-trim before-cursor))
+ (let* ((match (minuet-find-longest-match filtered-item before-cursor)))
+ (when (and match
+ (not (string-empty-p match))
+ (>= (length match) minuet-before-cursor-filter-length))
+ (setq filtered-item (substring filtered-item (length match))))))
+
+ ;; Filter against after-cursor context (trim from suffix)
+ (when (and after-cursor
+ (> minuet-after-cursor-filter-length 0))
+ (setq after-cursor (string-trim after-cursor))
+ (let* ((match (minuet-find-longest-match after-cursor filtered-item)))
+ (when (and match
+ (not (string-empty-p match))
+ (>= (length match) minuet-after-cursor-filter-length))
+ (setq filtered-item (substring filtered-item 0 (- (length
filtered-item) (length match)))))))
+
+ filtered-item))
+
+(cl-defun minuet-find-longest-match (a b)
+ "Find the longest string that is a prefix of A and a suffix of B.
+The function iterates from the longest possible match length downwards
+for efficiency. If A or B are not strings, it returns an empty
+string."
+ (unless (and (stringp a) (stringp b))
+ (cl-return-from minuet-find-longest-match ""))
+ (let* ((len-a (length a))
+ (len-b (length b))
+ (max-len (min len-a len-b)))
+ (cl-loop for i from max-len downto 1
+ for prefix-a = (substring a 0 i)
+ for suffix-b = (substring b (- i))
+ when (string= prefix-a suffix-b)
+ do (cl-return-from minuet-find-longest-match prefix-a))
""))
-(defun minuet--filter-text (text sequence)
- "Remove the SEQUENCE and the rest part from TEXT."
- (cond
- ((or (null sequence) (null text)) text)
- ((equal sequence "") text)
- (t
- (let ((start (string-match-p (regexp-quote sequence) text)))
- (if start
- (substring text 0 start)
- text)))))
-
-(defun minuet--filter-sequence-in-items (items sequence)
- "For each item in ITEMS, apply `minuet--filter-text' with SEQUENCE."
- (mapcar (lambda (x) (minuet--filter-text x sequence))
- items))
-
(defun minuet--filter-context-sequence-in-items (items context)
"Apply the filter sequence in each item in ITEMS.
The filter sequence is obtained from CONTEXT."
- (minuet--filter-sequence-in-items
- items (minuet--make-context-filter-sequence
- (plist-get context :after-cursor)
- minuet-after-cursor-filter-length)))
+ (cl-loop for item in items
+ for filtered-item = (minuet--filter-text item context)
+ when (and filtered-item
+ (not (string= filtered-item "")))
+ collect filtered-item))
+
(defun minuet--stream-decode (response get-text-fn)
"Decode the RESPONSE using GET-TEXT-FN."
