Deploy 302f7a5

app/pipeline/nodes/generate.py

@@ -66,16 +66,17 @@ ANSWERING RULES — follow all of them every time:
    say it, do not say it — not even if you "know" it from training data.
 3. READ ALL PASSAGES. An answer may be spread across multiple passages — a blog intro
    in [1], technical details in [3], project context in [5]. Synthesise all relevant
-   passages into one cohesive answer rather than stopping at the first match.
+   passages into one cohesive answer rather than stopping at the first match. Prioritise using varied sources (e.g., combining Resume with Project passages) to give a well-rounded answer.
 4. SCOPE. Use passages that directly address the question AND adjacent passages that
    provide supporting context, background, or related facts. If multiple passages
    contain information relevant to the query, you must cite all of them — do not
    cite only the first relevant passage and ignore others. A response about work
    experience that draws from one resume chunk must also cite any other resume chunk
    that adds detail.
-5. Cite every claim immediately after it with [N] where N is the passage number.
-   Example: "He optimised inference to 60 fps [1] by quantising the model [3]."
-   When a claim is backed by multiple passages, cite all: "He uses Python [1][4]."
+5. Cite at the end of the sentence or clause, not after every single item in a list.
+   Example: "He uses Python, Kotlin, and C++ [1][4]."
+   Do NOT cite like this: "He uses Python [1], Kotlin [1], and C++ [1]."
+   When a claim is backed by multiple passages, cite all: "[1][4]".
 6. If relevant passages contain limited facts, give a short answer covering exactly
    those facts — a short confident answer beats a padded hallucinated one.
 7. Vary your sentence openers. Never start two consecutive sentences with "Darshan".
@@ -324,7 +325,7 @@ def _normalise_answer_text(answer: str, max_citation_index: int) -> str:
 
     cleaned = _GEN_HTML_TAG_RE.sub("", answer)
     cleaned = re.sub(r"\[(\d+)\]", _keep_valid_citation, cleaned)
-    cleaned = re.sub(r"(\[\d+\])(\1)+", r"\1", cleaned)
+    cleaned = re.sub(r"(\[\d+\])(?:\s*\1)+", r"\1", cleaned)
     cleaned = re.sub(r"\s+([,.;:!?])", r"\1", cleaned)
     cleaned = re.sub(r"\n{3,}", "\n\n", cleaned)
     return cleaned.strip()

app/pipeline/nodes/retrieve.py

@@ -64,9 +64,7 @@ _FOCUS_KEYWORDS: dict[frozenset[str], str] = {
     frozenset({"experience", "work", "job", "role", "career", "internship",
                "skills", "skill", "education", "degree", "university",
                "certification", "certifications", "qualified", "resume", "cv",
-               "employment", "professional", "placement", "history",
-               "tech", "stack", "technology", "technologies", "framework",
-               "frameworks", "tool", "tools", "tooling", "language", "languages"}): "cv",
+               "employment", "professional", "placement", "history"}): "cv",
     frozenset({"project", "built", "build", "developed", "architecture",
                "system", "platform", "app", "application"}): "project",
     frozenset({"blog", "post", "article", "wrote", "writing", "published"}): "blog",