Wir stellen vor: Twilix – eine LangChain-Alternative für Retrieval-Augmented Generation | von Aaron | Juni 2023

0
30


Dieser Artikel soll die Lösung von Twilix im direkten Vergleich mit LangChain anhand eines von LangChain erstellten Beispiels vorstellen.

Das Folgende ist eine Beispielantwort aus dem Notion-QA-Beispiel von LangChain (https://github.com/hwchase17/notion-qa).

Die Lösung von Twilix konzentriert sich darauf, aus Sicht der Ergebnisleistung auf verschiedene Arten Überlegenheit zu bieten.

Verbesserte Era

python qa.py "How do you get a increase?"

Dies gibt Folgendes zurück:

Terminalantwort nach der Ausführung des obigen Terminalbefehls
Reply: To get a increase, one can convey up wage at any second they need 
and good causes to lift offcycle embody drastic change in position and
tasks, very steep studying curve which makes present pay unfair
for somebody's influence and capabilities, new data internally or
externally, and new data by benchmarking research or throughout hiring.
Not so good causes to provide a increase embody a colleague obtained one other supply
and makes use of this to get a increase, 'I earn X for this firm, that is a variety of
cash, I desire a piece of the pie', and colleague simply needs an incentive to
keep. When figuring out somebody's wage within the first place, Blendle weighs
up a number of variables together with benchmark, traditional figures, and shortage &
market worth. When life occurs, one can agree on what is required for this
individual on this particular scenario and when individuals work exhausting, they need to be
poked to take day without work so that they recharge. For longer leaves of absence, HR
must be consulted.

Sources: Notion_DB/Blendle's Worker Handbook a834d55573614857a48a9ce9ec4194e3/Wage Evaluation fd385c04dff04c2bb298f7085bf9e78d.md, Notion_DB/Blendle's Worker Handbook a834d55573614857a48a9ce9ec4194e3/Holidays x 5b552479e5ad40b49ce71fa68840ce35.md

Was ist additionally falsch an dieser Antwort?

  • Die Antwort hat folgende Probleme:
  • Das ist nicht der Fall Wirklich Für jemanden, der im Handbuch bei Blendle wörtlich fragt: „Wie kann jemand eine Gehaltserhöhung bekommen?“, ergibt das Sinn – warum kommt hier der Urlaubsurlaub zur Sprache?

Ich glaube, dass dies im Second ein umfassenderes Downside bei LangChain darstellt – es wird derzeit nicht viel Wert darauf gelegt, gute Ergebnisse zu erzielen, sondern nur auf die Bereitstellung hochwertiger Wrapper.

Ich wollte diese Lösung sofort mit Twilix testen. Ich habe denselben Splitter usw. verwendet, verlasse mich aber auf unsere interne proprietäre Lösung. Wenn Sie dies gegen LangChain ausführen, erhalten Sie Folgendes:

To get a increase at Blendle, there are just a few key components to think about.

First, it is vital to have motive for requesting a increase. Legitimate
causes embody a drastic change in position and tasks, a steep
studying curve that makes your present pay unfair, or new data both
internally or externally that justifies a wage improve. It is vital to
be capable of quantify and supply proof for these causes.

Nonetheless, there are additionally not-so-good causes to request a increase, akin to a
colleague getting one other supply and utilizing it as leverage, or just wanting
a bigger piece of the wage pie with none important adjustments in position or
efficiency. These causes are usually not thought of legitimate grounds for a
increase.

Along with having a sound motive, it is vital to think about
the general influence on the corporate. Earlier than deciding on a increase, it's best to ask
questions like:

- If we have been to rent somebody as we speak in your present job, what would we pay
them?
- What can be the fee to Blendle as a enterprise to discover a new individual
to switch you, contemplating components like shortage, time to rent, and the
influence on day-to-day operations?

These questions assist assess the worth of the increase in relation to Blendle as
an entire.

Lastly, it is price noting that the method of figuring out salaries at Blendle
takes under consideration benchmarking research and comparisons with inside friends.
It additionally considers traditional components like expertise, training, particular
data, and experience. Moreover, market worth and shortage are taken
into consideration.

For those who're not sure about how you can go about getting a increase or have another
questions, it is at all times finest to talk together with your lead and HR for steerage and
help.

Die Lösungen selbst entsprechen eher der Frage und liefern keine überflüssigen Antworten. Warum das? Die Lösung von Twilix zielt darauf ab, den Lärmpegel sofort zu minimieren. Der verrückte Teil? Es lässt sich auch problemlos in LangChain integrieren, sodass Sie die Ladeprogramme weiterhin verwenden und dennoch gute Lösungen erhalten können.

Verbessertes Abrufen

Schauen wir uns eine andere Frage an: „Gibt es Essen im Büro?“

LangChains Antwort:

Sure, there's lunch within the canteen at round 12 o'clock, freed from value. There 
are additionally snacks and drinks obtainable after All-hands shows. Nonetheless,
it's endorsed to convey your personal snacks if you're an enormous snack lover or
do not wish to be left empty-handed.

Twilix‘ Antwort:

Sure, there's meals within the workplace. Lunch is offered within the canteen round 
12 o'clock, freed from value. The aim is to have sufficient meals for everybody,
however generally they might run out of inventory. As well as, luxurious merchandise or
snacks are sometimes ordered to maintain issues particular and inside price range.
You probably have any food-related options, you may ship them to
officemanagement@blendle.com. Nonetheless, they can not cater to all particular person
wants.

Die Antwort von Twilix ist zwar ausführlicher, aber auch sachlich korrekter. Da diese beiden Antworten jedoch so unterschiedlich waren, musste ich das Originalquellenhandbuch durchsehen:

Die Antwort von Twilix enthielt diese Quelle.

~12 o' clock, there's lunch within the canteen, freed from value. 
Joëlle is in command of lunch — ask her in the event you want something (allergic reactions for
instance).

What you may anticipate from lunch:

- Wholesome: We attempt to supply merchandise that aren't dangerous for you!
- Sufficient for everybody: The aim is to have sufficient meals for everybody. However
generally we simply run out of inventory, dangerous luck... No worries, we are going to get some
extra subsequent time.
- Selection: Variation is sweet. However it's additionally good to maintain some issues particular.
Like luxurious merchandise or snacks, we order these sometimes. This could maintain
it particular and inside our price range.
- You probably have any meals associated options, be free to ship them to
officemanagement@blendle.com. However keep in mind this, we can't cater to all
particular person wants. However we at all times take options into consideration.

Die Lösung von LangChain enthielt diese Quelle.

**Snacks and extra**

There will likely be beer, soda and snacks after All-hands shows.

On another common day:

- We would have chips and so on, however op = op. We order snacks generally.
- This additionally goes for the cookies, fizzy drinks, beer or membership mate.
- Fruit is one thing we wish to present on a regular basis. For those who discover we run out
of fruit too quick and too typically, ping our Workplace Supervisor so she will be able to repair this.

In case you are an enormous snack lover or do not wish to be left empty-handed, think about
bringing your personal snacks.

Was ist additionally schief gelaufen und was sind die Unterschiede?

Bei der Durchführung dieser RAG-Lösung ist es wichtig, tatsächlich relevante Suchergebnisse in Ihre Lösung zu integrieren. Die Vektorsuche von LangChain hat in dieser State of affairs das Ergebnis „Snacks“ und nicht das nützlichere Ergebnis „Lebensmittel“ erfasst und daher keine bessere Antwort geliefert.

Es ist wichtig zu beachten, dass bei der Erstellung von RAG-Lösungen für die Abrufergebnisse häufig eine Fehlerquote erforderlich ist, wenn es sich um eine Vektorsuche handelt. Daher ist es sinnvoll, die Anzahl der Suchergebnisse zu erhöhen, die in die Eingabeaufforderung eingespeist werden. Für Entwickler, die Lösungen entwickeln möchten, die für Unternehmen nützlich sind, ist es wichtig, sich darauf zu konzentrieren, gute, relevante Suchergebnisse zu erhalten. Schlechtes RAG bedeutet Fehlinformationen und unvollständigen Kontext.

Zusammenfassung

Im Auslieferungszustand lässt Langchains RAG einiges zu wünschen übrig. Die RAG-Lösung von Twilix bietet umfassendere Antworten durch bessere Generierung und sofortigen Abruf – und das ist der Schwerpunkt dessen, was Twilix entwickelt. Wenn Sie daran interessiert sind, eine Various zu LangChain für Retrieval-Augmented Generaiton auszuprobieren, empfehle ich die Tutorials auf docs.twilix.io!



Source link

HINTERLASSEN SIE EINE ANTWORT

Please enter your comment!
Please enter your name here