From 8b70b749d353fcc66cedc7b79defa04dd5de3eaa Mon Sep 17 00:00:00 2001
From: Peter Werry <peter@getunblocked.com>
Date: Fri, 26 Jun 2026 14:54:11 -0700
Subject: [PATCH] Step 6: full-text search

---
 src/db.ts                  | 3 +++
 src/synthesis/prompt.ts    | 4 ++--
 src/validation/validate.ts | 2 ++
 3 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/src/db.ts b/src/db.ts
index e821e05..fe3dcfe 100644
--- a/src/db.ts
+++ b/src/db.ts
@@ -45,12 +45,15 @@ export async function ensureIndexes(): Promise<void> {
   await prs.createIndex({ state: 1 });
   await prs.createIndex({ merged_at: -1 });
   await prs.createIndex({ created_at: -1 });
+  // Full-text index over content, enabling {$text: {$search: ...}} topic queries.
+  await prs.createIndex({ title: "text", body: "text" });
 
   const issues = await getCollection("issues");
   await issues.createIndex({ repo: 1, number: 1 }, { unique: true });
   await issues.createIndex({ "user.login": 1 });
   await issues.createIndex({ state: 1 });
   await issues.createIndex({ created_at: -1 });
+  await issues.createIndex({ title: "text", body: "text" });
 
   const users = await getUsersCollection();
   await users.createIndex({ login: 1 }, { unique: true });
diff --git a/src/synthesis/prompt.ts b/src/synthesis/prompt.ts
index a8cac9a..b54df4a 100644
--- a/src/synthesis/prompt.ts
+++ b/src/synthesis/prompt.ts
@@ -9,8 +9,8 @@ ${schemaContext}
 - Use ONLY fields that appear in the schema above. Never invent fields.
 - Pick the single collection that best answers the question.
 - Allowed pipeline stages: $match, $sort, $limit, $count, $group, $project, $unwind.
-- Allowed filter operators: $eq, $ne, $gt, $gte, $lt, $lte, $in, $nin, $and, $or, $not, $exists, $size, $elemMatch.
-- Do NOT use $regex on free-text fields (title, body, comments.body). Match exact field values instead.
+- Allowed filter operators: $eq, $ne, $gt, $gte, $lt, $lte, $in, $nin, $and, $or, $not, $exists, $size, $elemMatch, $text.
+- For topic/keyword search over PR or issue content, use a text search: put {$text: {$search: "keywords"}} inside a $match (backed by a text index on title + body). Combine it with other filters in the SAME $match (author login, state, dates). Use a phrase like "\"code review\"" for exact phrases. Never use $regex on text fields. $text searches only title and body.
 - Timestamps are ISO-8601 strings; ISO strings sort lexicographically, so compare them directly with $gte/$lte.
 - For relative dates ("today", "last week", "past 3 days"), compute the cutoff from the "Current time" given in the user message — never guess the current date.
 - A PR is "merged" when merged is true (or merged_at is not null). "Open"/"closed" use the state field.
diff --git a/src/validation/validate.ts b/src/validation/validate.ts
index 2f4c552..b4776df 100644
--- a/src/validation/validate.ts
+++ b/src/validation/validate.ts
@@ -16,6 +16,8 @@ const ALLOWED_OPERATORS = new Set([
   // query operators
   "$eq", "$ne", "$gt", "$gte", "$lt", "$lte", "$in", "$nin",
   "$and", "$or", "$not", "$nor", "$exists", "$size", "$elemMatch",
+  // full-text search (backed by the title/body text index)
+  "$text", "$search", "$language", "$caseSensitive", "$diacriticSensitive",
   // accumulators / group expressions
   "$sum", "$avg", "$min", "$max", "$first", "$last", "$push", "$addToSet",
 ]);